Shu Kong

Neehar Peri, Achal Dave, Deva Ramanan et al.

Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for training 3D detectors, particularly on large-scale lidar data. Surprisingly, although semantic class labels naturally follow a long-tailed distribution, contemporary benchmarks focus on only a few common classes (e.g., pedestrian and car) and neglect many rare classes in-the-tail (e.g., debris and stroller). However, AVs must still detect rare classes to ensure safe operation. Moreover, semantic classes are often organized within a hierarchy, e.g., tail classes such as child and construction-worker are arguably subclasses of pedestrian. However, such hierarchical relationships are often ignored, which may lead to misleading estimates of performance and missed opportunities for algorithmic innovation. We address these challenges by formally studying the problem of Long-Tailed 3D Detection (LT3D), which evaluates on all classes, including those in-the-tail. We evaluate and innovate upon popular 3D detection codebases, such as CenterPoint and PointPillars, adapting them for LT3D. We develop hierarchical losses that promote feature sharing across common-vs-rare classes, as well as improved detection metrics that award partial credit to "reasonable" mistakes respecting the hierarchy (e.g., mistaking a child for an adult). Finally, we point out that fine-grained tail class accuracy is particularly improved via multimodal fusion of RGB images with LiDAR; simply put, small fine-grained classes are challenging to identify from sparse (lidar) geometry alone, suggesting that multimodal cues are crucial to long-tailed 3D detection. Our modifications improve accuracy by 5% AP on average for all classes, and dramatically improve AP for rare classes (e.g., stroller AP improves from 3.6 to 31.6)! Our code is available at https://github.com/neeharperi/LT3D

Zhiqiu Lin, Deepak Pathak, Yu-Xiong Wang et al. · cmu

Lifelong learners must recognize concept vocabularies that evolve over time. A common yet underexplored scenario is learning with class labels that continually refine/expand old classes. For example, humans learn to recognize ${\tt dog}$ before dog breeds. In practical settings, dataset $\textit{versioning}$ often introduces refinement to ontologies, such as autonomous vehicle benchmarks that refine a previous ${\tt vehicle}$ class into ${\tt school-bus}$ as autonomous operations expand to new cities. This paper formalizes a protocol for studying the problem of $\textit{Learning with Evolving Class Ontology}$ (LECO). LECO requires learning classifiers in distinct time periods (TPs); each TP introduces a new ontology of "fine" labels that refines old ontologies of "coarse" labels (e.g., dog breeds that refine the previous ${\tt dog}$). LECO explores such questions as whether to annotate new data or relabel the old, how to leverage coarse labels, and whether to finetune the previous TP's model or train from scratch. To answer these questions, we leverage insights from related problems such as class-incremental learning. We validate them under the LECO protocol through the lens of image classification (CIFAR and iNaturalist) and semantic segmentation (Mapillary). Our experiments lead to surprising conclusions; while the current status quo is to relabel existing datasets with new ontologies (such as COCO-to-LVIS or Mapillary1.2-to-2.0), LECO demonstrates that a far better strategy is to annotate $\textit{new}$ data with the new ontology. However, this produces an aggregate dataset with inconsistent old-vs-new labels, complicating learning. To address this challenge, we adopt methods from semi-supervised and partial-label learning. Such strategies can surprisingly be made near-optimal, approaching an "oracle" that learns on the aggregate dataset exhaustively labeled with the newest ontology.

Meng Wei, Xiaoyu Yue, Wenwei Zhang et al.

Segmenting and recognizing diverse object parts is a crucial ability in applications spanning various computer vision and robotic tasks. While significant progress has been made in object-level Open-Vocabulary Semantic Segmentation (OVSS), i.e., segmenting objects with arbitrary text, the corresponding part-level research poses additional challenges. Firstly, part segmentation inherently involves intricate boundaries, while limited annotated data compounds the challenge. Secondly, part segmentation introduces an open granularity challenge due to the diverse and often ambiguous definitions of parts in the open world. Furthermore, the large-scale vision and language models, which play a key role in the open vocabulary setting, struggle to recognize parts as effectively as objects. To comprehensively investigate and tackle these challenges, we propose an Open-Vocabulary Part Segmentation (OV-PARTS) benchmark. OV-PARTS includes refined versions of two publicly available datasets: Pascal-Part-116 and ADE20K-Part-234. And it covers three specific tasks: Generalized Zero-Shot Part Segmentation, Cross-Dataset Part Segmentation, and Few-Shot Part Segmentation, providing insights into analogical reasoning, open granularity and few-shot adapting abilities of models. Moreover, we analyze and adapt two prevailing paradigms of existing object-level OVSS methods for OV-PARTS. Extensive experimental analysis is conducted to inspire future research in leveraging foundational models for OV-PARTS. The code and dataset are available at https://github.com/OpenRobotLab/OV_PARTS.

Shubham Gupta, Jeet Kanjani, Mengtian Li et al. · gatech

We focus on the task of far-field 3D detection (Far3Det) of objects beyond a certain distance from an observer, e.g., $>$50m. Far3Det is particularly important for autonomous vehicles (AVs) operating at highway speeds, which require detections of far-field obstacles to ensure sufficient braking distances. However, contemporary AV benchmarks such as nuScenes underemphasize this problem because they evaluate performance only up to a certain distance (50m). One reason is that obtaining far-field 3D annotations is difficult, particularly for lidar sensors that produce very few point returns for far-away objects. Indeed, we find that almost 50% of far-field objects (beyond 50m) contain zero lidar points. Secondly, current metrics for 3D detection employ a "one-size-fits-all" philosophy, using the same tolerance thresholds for near and far objects, inconsistent with tolerances for both human vision and stereo disparities. Both factors lead to an incomplete analysis of the Far3Det task. For example, while conventional wisdom tells us that high-resolution RGB sensors should be vital for 3D detection of far-away objects, lidar-based methods still rank higher compared to RGB counterparts on the current benchmark leaderboards. As a first step towards a Far3Det benchmark, we develop a method to find well-annotated scenes from the nuScenes dataset and derive a well-annotated far-field validation set. We also propose a Far3Det evaluation protocol and explore various 3D detection methods for Far3Det. Our result convincingly justifies the long-held conventional wisdom that high-resolution RGB improves 3D detection in the far-field. We further propose a simple yet effective method that fuses detections from RGB and lidar detectors based on non-maximum suppression, which remarkably outperforms state-of-the-art 3D detectors in the far-field.

Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan et al.

In the real open world, data tends to follow long-tailed class distributions, motivating the well-studied long-tailed recognition (LTR) problem. Naive training produces models that are biased toward common classes in terms of higher accuracy. The key to addressing LTR is to balance various aspects including data distribution, training losses, and gradients in learning. We explore an orthogonal direction, weight balancing, motivated by the empirical observation that the naively trained classifier has "artificially" larger weights in norm for common classes (because there exists abundant data to train them, unlike the rare classes). We investigate three techniques to balance weights, L2-normalization, weight decay, and MaxNorm. We first point out that L2-normalization "perfectly" balances per-class weights to be unit norm, but such a hard constraint might prevent classes from learning better classifiers. In contrast, weight decay penalizes larger weights more heavily and so learns small balanced weights; the MaxNorm constraint encourages growing small weights within a norm ball but caps all the weights by the radius. Our extensive study shows that both help learn balanced weights and greatly improve the LTR accuracy. Surprisingly, weight decay, although underexplored in LTR, significantly improves over prior work. Therefore, we adopt a two-stage training paradigm and propose a simple approach to LTR: (1) learning features using the cross-entropy loss by tuning weight decay, and (2) learning classifiers using class-balanced loss by tuning weight decay and MaxNorm. Our approach achieves the state-of-the-art accuracy on five standard benchmarks, serving as a future baseline for long-tailed recognition.

Jia Shi, Gautam Gare, Jinjin Tian et al.

We tackle the challenge of predicting models' Out-of-Distribution (OOD) performance using in-distribution (ID) measurements without requiring OOD data. Existing evaluations with "Effective Robustness", which use ID accuracy as an indicator of OOD accuracy, encounter limitations when models are trained with diverse supervision and distributions, such as class labels (Vision Models, VMs, on ImageNet) and textual descriptions (Visual-Language Models, VLMs, on LAION). VLMs often generalize better to OOD data than VMs despite having similar or lower ID performance. To improve the prediction of models' OOD performance from ID measurements, we introduce the Lowest Common Ancestor (LCA)-on-the-Line framework. This approach revisits the established concept of LCA distance, which measures the hierarchical distance between labels and predictions within a predefined class hierarchy, such as WordNet. We assess 75 models using ImageNet as the ID dataset and five significantly shifted OOD variants, uncovering a strong linear correlation between ID LCA distance and OOD top-1 accuracy. Our method provides a compelling alternative for understanding why VLMs tend to generalize better. Additionally, we propose a technique to construct a taxonomic hierarchy on any dataset using K-means clustering, demonstrating that LCA distance is robust to the constructed taxonomic hierarchy. Moreover, we demonstrate that aligning model predictions with class taxonomies, through soft labels or prompt engineering, can enhance model generalization. Open source code in our Project Page: https://elvishelvis.github.io/papers/lca/.

Samia Shafique, Bailey Kong, Shu Kong et al.

Shoe tread impressions are one of the most common types of evidence left at crime scenes. However, the utility of such evidence is limited by the lack of databases of footwear prints that cover the large and growing number of distinct shoe models. Moreover, the database is preferred to contain the 3D shape, or depth, of shoe-tread photos so as to allow for extracting shoeprints to match a query (crime-scene) print. We propose to address this gap by leveraging shoe-tread photos collected by online retailers. The core challenge is to predict depth maps for these photos. As they do not have ground-truth 3D shapes allowing for training depth predictors, we exploit synthetic data that does. We develop a method termed ShoeRinsics that learns to predict depth by leveraging a mix of fully supervised synthetic data and unsupervised retail image data. In particular, we find domain adaptation and intrinsic image decomposition techniques effectively mitigate the synthetic-real domain gap and yield significantly better depth prediction. To validate our method, we introduce 2 validation sets consisting of shoe-tread image and print pairs and define a benchmarking protocol to quantify the quality of predicted depth. On this benchmark, ShoeRinsics outperforms existing methods of depth prediction and synthetic-to-real domain adaptation.

Anirudh S Chakravarthy, Meghana Reddy Ganesina, Peiyun Hu et al.

Addressing Lidar Panoptic Segmentation (LPS ) is crucial for safe deployment of autonomous vehicles. LPS aims to recognize and segment lidar points w.r.t. a pre-defined vocabulary of semantic classes, including thing classes of countable objects (e.g., pedestrians and vehicles) and stuff classes of amorphous regions (e.g., vegetation and road). Importantly, LPS requires segmenting individual thing instances (e.g., every single vehicle). Current LPS methods make an unrealistic assumption that the semantic class vocabulary is fixed in the real open world, but in fact, class ontologies usually evolve over time as robots encounter instances of novel classes that are considered to be unknowns w.r.t. the pre-defined class vocabulary. To address this unrealistic assumption, we study LPS in the Open World (LiPSOW): we train models on a dataset with a pre-defined semantic class vocabulary and study their generalization to a larger dataset where novel instances of thing and stuff classes can appear. This experimental setting leads to interesting conclusions. While prior art train class-specific instance segmentation methods and obtain state-of-the-art results on known classes, methods based on class-agnostic bottom-up grouping perform favorably on classes outside of the initial class vocabulary (i.e., unknown classes). Unfortunately, these methods do not perform on-par with fully data-driven methods on known classes. Our work suggests a middle ground: we perform class-agnostic point clustering and over-segment the input cloud in a hierarchical fashion, followed by binary point segment classification, akin to Region Proposal Network [1]. We obtain the final point cloud segmentation by computing a cut in the weighted hierarchical tree of point segments, independently of semantic classification. Remarkably, this unified approach leads to strong performance on both known and unknown classes.

Shubham Parashar, Zhiqiu Lin, Yanan Li et al.

Trained on web-scale image-text pairs, Vision-Language Models (VLMs) such as CLIP can recognize images of common objects in a zero-shot fashion. However, it is underexplored how to use CLIP for zero-shot recognition of highly specialized concepts, e.g., species of birds, plants, and animals, for which their scientific names are written in Latin or Greek. Indeed, CLIP performs poorly for zero-shot species recognition with prompts that use scientific names, e.g., "a photo of Lepus Timidus" (which is a scientific name in Latin). Because these names are usually not included in CLIP's training set. To improve performance, prior works propose to use large-language models (LLMs) to generate descriptions (e.g., of species color and shape) and additionally use them in prompts. We find that they bring only marginal gains. Differently, we are motivated to translate scientific names (e.g., Lepus Timidus) to common English names (e.g., mountain hare) and use such in the prompts. We find that common names are more likely to be included in CLIP's training set, and prompting them achieves 2$\sim$5 times higher accuracy on benchmarking datasets of fine-grained species recognition.

Tian Liu, Anwesha Basu, James Caverlee et al.

Semi-supervised few-shot learning (SSFSL) formulates real-world applications like ''auto-annotation'', as it aims to learn a model over a few labeled and abundant unlabeled examples to annotate the unlabeled ones. Despite the availability of powerful open-source Vision-Language Models (VLMs) and their pretraining data, the SSFSL literature largely neglects these open-source resources. In contrast, the related area few-shot learning (FSL) has already exploited them to boost performance. Arguably, to achieve auto-annotation in the real world, SSFSL should leverage such open-source resources. To this end, we start by applying established SSL methods to finetune a VLM. Counterintuitively, they significantly underperform FSL baselines. Our in-depth analysis reveals the root cause: VLMs produce rather ''flat'' distributions of softmax probabilities. This results in zero utilization of unlabeled data and weak supervision signals. We address this issue with embarrassingly simple techniques: classifier initialization and temperature tuning. They jointly increase the confidence scores of pseudo-labels, improving the utilization rate of unlabeled data, and strengthening supervision signals. Building on this, we propose: Stage-Wise Finetuning with Temperature Tuning (SWIFT), which enables existing SSL methods to effectively finetune a VLM on limited labeled data, abundant unlabeled data, and task-relevant but noisy data retrieved from the VLM's pretraining set. Extensive experiments on five SSFSL benchmarks show that SWIFT outperforms recent FSL and SSL methods by $\sim$5 accuracy points. SWIFT even rivals supervised learning, which finetunes VLMs with the unlabeled data being labeled with ground truth!

Anish Madan, Neehar Peri, Shu Kong et al.

The era of vision-language models (VLMs) trained on web-scale datasets challenges conventional formulations of "open-world" perception. In this work, we revisit the task of few-shot object detection (FSOD) in the context of recent foundational VLMs. First, we point out that zero-shot predictions from VLMs such as GroundingDINO significantly outperform state-of-the-art few-shot detectors (48 vs. 33 AP) on COCO. Despite their strong zero-shot performance, such foundation models may still be sub-optimal. For example, trucks on the web may be defined differently from trucks for a target application such as autonomous vehicle perception. We argue that the task of few-shot recognition can be reformulated as aligning foundation models to target concepts using a few examples. Interestingly, such examples can be multi-modal, using both text and visual cues, mimicking instructions that are often given to human annotators when defining a target concept of interest. Concretely, we propose Foundational FSOD, a new benchmark protocol that evaluates detectors pre-trained on any external data and fine-tuned on multi-modal (text and visual) K-shot examples per target class. We repurpose nuImages for Foundational FSOD, benchmark several popular open-source VLMs, and provide an empirical analysis of state-of-the-art methods. Lastly, we discuss our recent CVPR 2024 Foundational FSOD competition and share insights from the community. Notably, the winning team significantly outperforms our baseline by 23.3 mAP! Our code and dataset splits are available at https://github.com/anishmadan23/foundational_fsod

Qianqian Shen, Yunhan Zhao, Nahyun Kwon et al.

Instance detection (InsDet) is a long-lasting problem in robotics and computer vision, aiming to detect object instances (predefined by some visual examples) in a cluttered scene. Despite its practical significance, its advancement is overshadowed by Object Detection, which aims to detect objects belonging to some predefined classes. One major reason is that current InsDet datasets are too small in scale by today's standards. For example, the popular InsDet dataset GMU (published in 2016) has only 23 instances, far less than COCO (80 classes), a well-known object detection dataset published in 2014. We are motivated to introduce a new InsDet dataset and protocol. First, we define a realistic setup for InsDet: training data consists of multi-view instance captures, along with diverse scene images allowing synthesizing training images by pasting instance images on them with free box annotations. Second, we release a real-world database, which contains multi-view capture of 100 object instances, and high-resolution (6k x 8k) testing images. Third, we extensively study baseline methods for InsDet on our dataset, analyze their performance and suggest future work. Somewhat surprisingly, using the off-the-shelf class-agnostic segmentation model (Segment Anything Model, SAM) and the self-supervised feature representation DINOv2 performs the best, achieving >10 AP better than end-to-end trained InsDet models that repurpose object detectors (e.g., FasterRCNN and RetinaNet).

Nahyun Kwon, Qian Lu, Muhammad Hasham Qazi et al.

In our increasingly diverse society, everyday physical interfaces often present barriers, impacting individuals across various contexts. This oversight, from small cabinet knobs to identical wall switches that can pose different contextual challenges, highlights an imperative need for solutions. Leveraging low-cost 3D-printed augmentations such as knob magnifiers and tactile labels seems promising, yet the process of discovering unrecognized barriers remains challenging because disability is context-dependent. We introduce AccessLens, an end-to-end system designed to identify inaccessible interfaces in daily objects, and recommend 3D-printable augmentations for accessibility enhancement. Our approach involves training a detector using the novel AccessDB dataset designed to automatically recognize 21 distinct Inaccessibility Classes (e.g., bar-small and round-rotate) within 6 common object categories (e.g., handle and knob). AccessMeta serves as a robust way to build a comprehensive dictionary linking these accessibility classes to open-source 3D augmentation designs. Experiments demonstrate our detector's performance in detecting inaccessible objects.

Yechi Ma, Neehar Peri, Achal Dave et al.

Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for training 3D detectors. While class labels naturally follow a long-tailed distribution in the real world, existing benchmarks only focus on a few common classes (e.g., pedestrian and car) and neglect many rare but crucial classes (e.g., emergency vehicle and stroller). However, AVs must reliably detect both common and rare classes for safe operation in the open world. We address this challenge by formally studying the problem of Long-Tailed 3D Detection (LT3D), which evaluates all annotated classes, including those in-the-tail. We address LT3D with hierarchical losses that promote feature sharing across classes, and introduce diagnostic metrics that award partial credit to "reasonable" mistakes with respect to the semantic hierarchy. Further, we point out that rare-class accuracy is particularly improved via multi-modal late fusion (MMLF) of independently trained uni-modal LiDAR and RGB detectors. Such an MMLF framework allows us to leverage large-scale uni-modal datasets (with more examples for rare classes) to train better uni-modal detectors. Finally, we examine three critical components of our simple MMLF approach from first principles: whether to train 2D or 3D RGB detectors for fusion, whether to match RGB and LiDAR detections in 3D or the projected 2D image plane, and how to fuse matched detections. Extensive experiments reveal that 2D RGB detectors achieve better recognition accuracy for rare classes than 3D RGB detectors, matching on the 2D image plane mitigates depth estimation errors for better matching, and score calibration and probabilistic fusion notably improves the final performance further. Our MMLF significantly outperforms prior work for LT3D, particularly improving on the six rarest classes from 12.8 to 20.0 mAP! Our code and models are available on our project page.

Tong Wang, Yunhan Zhao, Shu Kong

Composed Image Retrieval (CIR) is the task of retrieving a target image from a database using a multimodal query, which consists of a reference image and a modification text. The text specifies how to alter the reference image to form a ''mental image'', based on which CIR should find the target image in the database. The fundamental challenge of CIR is that this ''mental image'' is not physically available and is only implicitly defined by the query. The contemporary literature pursues zero-shot methods and uses a Large Multimodal Model (LMM) to generate a textual description for a given multimodal query, and then employs a Vision-Language Model (VLM) for textual-visual matching to search for the target image. In contrast, we address CIR from first principles by directly generating the ''mental image'' for more accurate matching. Particularly, we prompt an LMM to generate a ''mental image'' for a given multimodal query and propose to use this ''mental image'' to search for the target image. As the ''mental image'' has a synthetic-to-real domain gap with real images, we also generate a synthetic counterpart for each real image in the database to facilitate matching. In this sense, our method uses LMM to construct a ``paracosm'', where it matches the multimodal query and database images. Hence, we call this method Paracosm. Notably, Paracosm is a training-free zero-shot CIR method. It significantly outperforms existing zero-shot methods on challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Zeyi Sun, Ye Fang, Tong Wu et al.

Contrastive Language-Image Pre-training (CLIP) plays an essential role in extracting valuable content information from images across diverse tasks. It aligns textual and visual modalities to comprehend the entire image, including all the details, even those irrelevant to specific tasks. However, for a finer understanding and controlled editing of images, it becomes crucial to focus on specific regions of interest, which can be indicated as points, masks, or boxes by humans or perception models. To fulfill the requirements, we introduce Alpha-CLIP, an enhanced version of CLIP with an auxiliary alpha channel to suggest attentive regions and fine-tuned with constructed millions of RGBA region-text pairs. Alpha-CLIP not only preserves the visual recognition ability of CLIP but also enables precise control over the emphasis of image contents. It demonstrates effectiveness in various tasks, including but not limited to open-world recognition, multimodal large language models, and conditional 2D / 3D generation. It has a strong potential to serve as a versatile tool for image-related tasks.

Yechi Ma, Wei Hua, Shu Kong

A crucial yet under-appreciated prerequisite in machine learning solutions for real-applications is data annotation: human annotators are hired to manually label data according to detailed, expert-crafted guidelines. This is often a laborious, tedious, and costly process. To study methods for facilitating data annotation, we introduce a new benchmark AnnoGuide: Auto-Annotation from Annotation Guidelines. It aims to evaluate automated methods for data annotation directly from expert-defined annotation guidelines, eliminating the need for manual labeling. As a case study, we repurpose the well-established nuScenes dataset, commonly used in autonomous driving research, which provides comprehensive annotation guidelines for labeling LiDAR point clouds with 3D cuboids across 18 object classes. These guidelines include a few visual examples and textual descriptions, but no labeled 3D cuboids in LiDAR data, making this a novel task of multi-modal few-shot 3D detection without 3D annotations. The advances of powerful foundation models (FMs) make AnnoGuide especially timely, as FMs offer promising tools to tackle its challenges. We employ a conceptually straightforward pipeline that (1) utilizes open-source FMs for object detection and segmentation in RGB images, (2) projects 2D detections into 3D using known camera poses, and (3) clusters LiDAR points within the frustum of each 2D detection to generate a 3D cuboid. Starting with a non-learned solution that leverages off-the-shelf FMs, we progressively refine key components and achieve significant performance improvements, boosting 3D detection mAP from 12.1 to 21.9! Nevertheless, our results highlight that AnnoGuide remains an open and challenging problem, underscoring the urgent need for developing LiDAR-based FMs. We release our code and models at GitHub: https://annoguide.github.io/annoguide3Dbenchmark

Tong Wang, Jiaqi Wang, Shu Kong

Pretrained on web-scale open data, VLMs offer powerful capabilities for solving downstream tasks after being adapted to task-specific labeled data. Yet, data labeling can be expensive and may demand domain expertise. Active Learning (AL) aims to reduce this expense by strategically selecting the most informative data for labeling and model training. Recent AL methods have explored VLMs but have not leveraged publicly available open data, such as VLM's pretraining data. In this work, we leverage such data by retrieving task-relevant examples to augment the task-specific examples. As expected, incorporating them significantly improves AL. Given that our method exploits open-source VLM and open data, we refer to it as Active Learning with Open Resources (ALOR). Additionally, most VLM-based AL methods use prompt tuning (PT) for model adaptation, likely due to its ability to directly utilize pretrained parameters and the assumption that doing so reduces the risk of overfitting to limited labeled data. We rigorously compare popular adaptation approaches, including linear probing (LP), finetuning (FT), and contrastive tuning (CT). We reveal two key findings: (1) All adaptation approaches benefit from incorporating retrieved data, and (2) CT resoundingly outperforms other approaches across AL methods. Further analysis of retrieved data reveals a naturally imbalanced distribution of task-relevant classes, exposing inherent biases within the VLM. This motivates our novel Tail First Sampling (TFS) strategy for AL, an embarrassingly simple yet effective method that prioritizes sampling data from underrepresented classes to label. Extensive experiments demonstrate that our final method, contrastively finetuning VLM on both retrieved and TFS-selected labeled data, significantly outperforms existing methods.

Yuzhu Wang, Lechao Cheng, Manni Duan et al.

Knowledge distillation (KD) exploits a large well-trained model (i.e., teacher) to train a small student model on the same dataset for the same task. Treating teacher features as knowledge, prevailing methods of knowledge distillation train student by aligning its features with the teacher's, e.g., by minimizing the KL-divergence between their logits or L2 distance between their intermediate features. While it is natural to believe that better alignment of student features to the teacher better distills teacher knowledge, simply forcing this alignment does not directly contribute to the student's performance, e.g., classification accuracy. In this work, we propose to align student features with class-mean of teacher features, where class-mean naturally serves as a strong classifier. To this end, we explore baseline techniques such as adopting the cosine distance based loss to encourage the similarity between student features and their corresponding class-means of the teacher. Moreover, we train the student to produce large-norm features, inspired by other lines of work (e.g., model pruning and domain adaptation), which find the large-norm features to be more significant. Finally, we propose a rather simple loss term (dubbed ND loss) to simultaneously (1) encourage student to produce large-\emph{norm} features, and (2) align the \emph{direction} of student features and teacher class-means. Experiments on standard benchmarks demonstrate that our explored techniques help existing KD methods achieve better performance, i.e., higher classification accuracy on ImageNet and CIFAR100 datasets, and higher detection precision on COCO dataset. Importantly, our proposed ND loss helps the most, leading to the state-of-the-art performance on these benchmarks. The source code is available at \url{https://github.com/WangYZ1608/Knowledge-Distillation-via-ND}.

Shubham Parashar, Zhiqiu Lin, Tian Liu et al.

Vision-language models (VLMs) excel in zero-shot recognition but their performance varies greatly across different visual concepts. For example, although CLIP achieves impressive accuracy on ImageNet (60-80%), its performance drops below 10% for more than ten concepts like night snake, presumably due to their limited presence in the pretraining data. However, measuring the frequency of concepts in VLMs' large-scale datasets is challenging. We address this by using large language models (LLMs) to count the number of pretraining texts that contain synonyms of these concepts. Our analysis confirms that popular datasets, such as LAION, exhibit a long-tailed concept distribution, yielding biased performance in VLMs. We also find that downstream applications of VLMs, including visual chatbots (e.g., GPT-4V) and text-to-image models (e.g., Stable Diffusion), often fail to recognize or generate images of rare concepts identified by our method. To mitigate the imbalanced performance of zero-shot VLMs, we propose REtrieval-Augmented Learning (REAL). First, instead of prompting VLMs using the original class names, REAL uses their most frequent synonyms found in pretraining texts. This simple change already outperforms costly human-engineered and LLM-enriched prompts over nine benchmark datasets. Second, REAL trains a linear classifier on a small yet balanced set of pretraining data retrieved using concept synonyms. REAL surpasses the previous zero-shot SOTA, using 400x less storage and 10,000x less training time!

Yunhan Zhao, Haoyu Ma, Shu Kong et al. · meta-ai

Egocentric sensors such as AR/VR devices capture human-object interactions and offer the potential to provide task-assistance by recalling 3D locations of objects of interest in the surrounding environment. This capability requires instance tracking in real-world 3D scenes from egocentric videos (IT3DEgo). We explore this problem by first introducing a new benchmark dataset, consisting of RGB and depth videos, per-frame camera pose, and instance-level annotations in both 2D camera and 3D world coordinates. We present an evaluation protocol which evaluates tracking performance in 3D coordinates with two settings for enrolling instances to track: (1) single-view online enrollment where an instance is specified on-the-fly based on the human wearer's interactions. and (2) multi-view pre-enrollment where images of an instance to be tracked are stored in memory ahead of time. To address IT3DEgo, we first re-purpose methods from relevant areas, e.g., single object tracking (SOT) -- running SOT methods to track instances in 2D frames and lifting them to 3D using camera pose and depth. We also present a simple method that leverages pretrained segmentation and detection models to generate proposals from RGB frames and match proposals with enrolled instance images. Our experiments show that our method (with no finetuning) significantly outperforms SOT-based approaches in the egocentric setting. We conclude by arguing that the problem of egocentric instance tracking is made easier by leveraging camera pose and using a 3D allocentric (world) coordinate representation.

Xiaogang Xu, Shu Kong, Tao Hu et al.

Pre-trained models with large-scale training data, such as CLIP and Stable Diffusion, have demonstrated remarkable performance in various high-level computer vision tasks such as image understanding and generation from language descriptions. Yet, their potential for low-level tasks such as image restoration remains relatively unexplored. In this paper, we explore such models to enhance image restoration. As off-the-shelf features (OSF) from pre-trained models do not directly serve image restoration, we propose to learn an additional lightweight module called Pre-Train-Guided Refinement Module (PTG-RM) to refine restoration results of a target restoration network with OSF. PTG-RM consists of two components, Pre-Train-Guided Spatial-Varying Enhancement (PTG-SVE), and Pre-Train-Guided Channel-Spatial Attention (PTG-CSA). PTG-SVE enables optimal short- and long-range neural operations, while PTG-CSA enhances spatial-channel attention for restoration-related learning. Extensive experiments demonstrate that PTG-RM, with its compact size ($<$1M parameters), effectively enhances restoration performance of various models across different tasks, including low-light enhancement, deraining, deblurring, and denoising.

Qianqian Shen, Yunhan Zhao, Nahyun Kwon et al.

Instance detection (InsDet) aims to localize specific object instances within a novel scene imagery based on given visual references. Technically, it requires proposal detection to identify all possible object instances, followed by instance-level matching to pinpoint the ones of interest. Its open-world nature supports its broad applications from robotics to AR/VR but also presents significant challenges: methods must generalize to unknown testing data distributions because (1) the testing scene imagery is unseen during training, and (2) there are domain gaps between visual references and detected proposals. Existing methods tackle these challenges by synthesizing diverse training examples or utilizing off-the-shelf foundation models (FMs). However, they only partially capitalize the available open-world information. In contrast, we approach InsDet from an Open-World perspective, introducing our method IDOW. We find that, while pretrained FMs yield high recall in instance detection, they are not specifically optimized for instance-level feature matching. Therefore, we adapt pretrained FMs for improved instance-level matching using open-world data. Our approach incorporates metric learning along with novel data augmentations, which sample distractors as negative examples and synthesize novel-view instances to enrich the visual references. Extensive experiments demonstrate that our method significantly outperforms prior works, achieving >10 AP over previous results on two recently released challenging benchmark datasets in both conventional and novel instance detection settings.

Samia Shafique, Shu Kong, Charless Fowlkes

Shoeprints are a common type of evidence found at crime scenes and are used regularly in forensic investigations. However, existing methods cannot effectively employ deep learning techniques to match noisy and occluded crime-scene shoeprints to a shoe database due to a lack of training data. Moreover, all existing methods match crime-scene shoeprints to clean reference prints, yet our analysis shows matching to more informative tread depth maps yields better retrieval results. The matching task is further complicated by the necessity to identify similarities only in corresponding regions (heels, toes, etc) of prints and shoe treads. To overcome these challenges, we leverage shoe tread images from online retailers and utilize an off-the-shelf predictor to estimate depth maps and clean prints. Our method, named CriSp, matches crime-scene shoeprints to tread depth maps by training on this data. CriSp incorporates data augmentation to simulate crime-scene shoeprints, an encoder to learn spatially-aware features, and a masking module to ensure only visible regions of crime-scene prints affect retrieval results. To validate our approach, we introduce two validation sets by reprocessing existing datasets of crime-scene shoeprints and establish a benchmarking protocol for comparison. On this benchmark, CriSp significantly outperforms state-of-the-art methods in both automated shoeprint matching and image retrieval tailored to this task.

Nicholas Milef, John Keyser, Shu Kong

Optical motion capture (mocap) requires accurately reconstructing the human body from retroreflective markers, including pose and shape. In a typical mocap setting, marker labeling is an important but tedious and error-prone step. Previous work has shown that marker labeling can be automated by using a structured template defining specific marker placements, but this places additional recording constraints. We propose to relax these constraints and solve for Unstructured Unlabeled Optical (UUO) mocap. Compared to the typical mocap setting that either labels markers or places them w.r.t a structured layout, markers in UUO mocap can be placed anywhere on the body and even on one specific limb (e.g., right leg for biomechanics research), hence it is of more practical significance. It is also more challenging. To solve UUO mocap, we exploit a monocular video captured by a single RGB camera, which does not require camera calibration. On this video, we run an off-the-shelf method to reconstruct and track a human individual, giving strong visual priors of human body pose and shape. With both the video and UUO markers, we propose an optimization pipeline towards marker identification, marker labeling, human pose estimation, and human body reconstruction. Our technical novelties include multiple hypothesis testing to optimize global orientation, and marker localization and marker-part matching to better optimize for body surface. We conduct extensive experiments to quantitatively compare our method against state-of-the-art approaches, including marker-only mocap and video-only human body/shape reconstruction. Experiments demonstrate that our method resoundingly outperforms existing methods on three established benchmark datasets for both full-body and partial-body reconstruction.

Kaijie Yin, Zhiyuan Zhang, Shu Kong et al.

In this paper, we propose Binarized Change Detection (BiCD), the first binary neural network (BNN) designed specifically for change detection. Conventional network binarization approaches, which directly quantize both weights and activations in change detection models, severely limit the network's ability to represent input data and distinguish between changed and unchanged regions. This results in significantly lower detection accuracy compared to real-valued networks. To overcome these challenges, BiCD enhances both the representational power and feature separability of BNNs, improving detection performance. Specifically, we introduce an auxiliary objective based on the Information Bottleneck (IB) principle, guiding the encoder to retain essential input information while promoting better feature discrimination. Since directly computing mutual information under the IB principle is intractable, we design a compact, learnable auxiliary module as an approximation target, leading to a simple yet effective optimization strategy that minimizes both reconstruction loss and standard change detection loss. Extensive experiments on street-view and remote sensing datasets demonstrate that BiCD establishes a new benchmark for BNN-based change detection, achieving state-of-the-art performance in this domain.

Sophie Zhou, Shu Kong

Art plagiarism detection plays a crucial role in protecting artists' copyrights and intellectual property, yet it remains a challenging problem in forensic analysis. In this paper, we address the task of recognizing plagiarized paintings and explaining the detected plagarisms by retrieving visually similar authentic artworks. To support this study, we construct a dataset by collecting painting photos and synthesizing plagiarized versions using generative AI, tailored to specific artists' styles. We first establish a baseline approach using off-the-shelf features from the visual foundation model DINOv2 to retrieve the most similar images in the database and classify plagiarism based on a similarity threshold. Surprisingly, this non-learned method achieves a high recognition accuracy of 97.2\% but suffers from low retrieval precision 29.0\% average precision (AP). To improve retrieval quality, we finetune DINOv2 with a metric learning loss using positive and negative sample pairs sampled in the database. The finetuned model greatly improves retrieval performance by 12\% AP over the baseline, though it unexpectedly results in a lower recognition accuracy (92.7\%). We conclude with insightful discussions and outline directions for future research.

Yuzhu Wang, Manni Duan, Shu Kong

Visual Prompt Tuning (VPT) is a parameter-efficient fune-tuning technique that adapts a pre-trained vision Transformer (ViT) by learning a small set of parameters in the input space, known as prompts. In VPT, we uncover ``burstiness'' in the values arising from the interaction of image patch embeddings, and the key and query projectors within Transformer's self-attention module. Furthermore, the values of patch embeddings and the key and query projectors exhibit Laplacian and hyper-Laplacian distribution, respectively. Intuitively, these non-Gaussian distributions pose challenges for learning prompts. To address this, we propose whitening these data, de-correlating them and equalizing their variance towards more Gaussian before learning prompts. We derive the whitening matrix over random image patch embeddings and ViT's key and query projectors, and multiply it with the prompt to be learned in a bilinear manner. Surprisingly, this method significantly accelerates prompt tuning and boosts accuracy, e.g., $>$25 accuracy points on the CUB dataset; interestingly, it learns ``bursty prompts''. Extending the bilinear model which is known to introduce burstiness, we present a compact, low-rank version by learning two smaller matrices whose multiplication yields the final prompts. We call the proposed methods Bilinear Prompt Tuning (BPT). Extensive experiments across multiple benchmark datasets demonstrate that BPT methods not only outperform various VPT methods but also reduce parameter count and computation overhead.

Hanxin Wang, Tian Liu, Shu Kong

Pretrained VLMs achieve strong performance on downstream tasks when adapted with just a few labeled examples. As the adapted models inevitably encounter out-of-distribution (OOD) test data that deviates from the in-distribution (ID) task-specific training data, enhancing OOD generalization in few-shot adaptation is critically important. We study robust few-shot VLM adaptation, aiming to increase both ID and OOD accuracy. By comparing different adaptation methods (e.g., prompt tuning, linear probing, contrastive finetuning, and full finetuning), we uncover three key findings: (1) finetuning with proper hyperparameters significantly outperforms the popular VLM adaptation methods prompt tuning and linear probing; (2) visual encoder-only finetuning achieves better efficiency and accuracy than contrastively finetuning both visual and textual encoders; (3) finetuning the top layers of the visual encoder provides the best balance between ID and OOD accuracy. Building on these findings, we propose partial finetuning of the visual encoder empowered with two simple augmentation techniques: (1) retrieval augmentation which retrieves task-relevant data from the VLM's pretraining dataset to enhance adaptation, and (2) adversarial perturbation which promotes robustness during finetuning. Results show that the former/latter boosts OOD/ID accuracy while slightly sacrificing the ID/OOD accuracy. Yet, perhaps understandably, naively combining the two does not maintain their best OOD/ID accuracy. We address this dilemma with the developed SRAPF, Stage-wise Retrieval Augmentation-based Adversarial Partial Finetuning. SRAPF consists of two stages: (1) partial finetuning the visual encoder using both ID and retrieved data, and (2) adversarial partial finetuning with few-shot ID data. Extensive experiments demonstrate that SRAPF achieves the state-of-the-art ID and OOD accuracy on the ImageNet OOD benchmarks.

Tian Liu, Huixin Zhang, Shubham Parashar et al.

Few-shot recognition (FSR) aims to train a classification model with only a few labeled examples of each concept concerned by a downstream task, where data annotation cost can be prohibitively high. We develop methods to solve FSR by leveraging a pretrained Vision-Language Model (VLM). We particularly explore retrieval-augmented learning (RAL), which retrieves open data, e.g., the VLM's pretraining dataset, to learn models for better serving downstream tasks. RAL has been studied in zero-shot recognition but remains under-explored in FSR. Although applying RAL to FSR may seem straightforward, we observe interesting and novel challenges and opportunities. First, somewhat surprisingly, finetuning a VLM on a large amount of retrieved data underperforms state-of-the-art zero-shot methods. This is due to the imbalanced distribution of retrieved data and its domain gaps with the few-shot examples in the downstream task. Second, more surprisingly, we find that simply finetuning a VLM solely on few-shot examples significantly outperforms previous FSR methods, and finetuning on the mix of retrieved and few-shot data yields even better results. Third, to mitigate the imbalanced distribution and domain gap issues, we propose Stage-Wise retrieval-Augmented fineTuning (SWAT), which involves end-to-end finetuning on mixed data in the first stage and retraining the classifier on the few-shot data in the second stage. Extensive experiments on nine popular benchmarks demonstrate that SWAT significantly outperforms previous methods by >6% accuracy.

Yechi Ma, Yanan Li, Wei Hua et al.

Roadside monocular 3D detection requires detecting objects of predefined classes in an RGB frame and predicting their 3D attributes, such as bird's-eye-view (BEV) locations. It has broad applications in traffic control, vehicle-vehicle communication, and vehicle-infrastructure cooperative perception. To address this task, we introduce Promptable 3D Detector (Pro3D), a novel detector design that leverages 2D detections as prompts. We build our Pro3D upon two key insights. First, compared to a typical 3D detector, a 2D detector is ``easier'' to train due to fewer loss terms and performs significantly better at localizing objects w.r.t 2D metrics. Second, once 2D detections precisely locate objects in the image, a 3D detector can focus on lifting these detections into 3D BEV, especially when fixed camera pose or scene geometry provide an informative prior. To encode and incorporate 2D detections, we explore three methods: (a) concatenating features from both 2D and 3D detectors, (b) attentively fusing 2D and 3D detector features, and (c) encoding properties of predicted 2D bounding boxes \{$x$, $y$, width, height, label\} and attentively fusing them with the 3D detector feature. Interestingly, the third method significantly outperforms the others, underscoring the effectiveness of 2D detections as prompts that offer precise object targets and allow the 3D detector to focus on lifting them into 3D. Pro3D is adaptable for use with a wide range of 2D and 3D detectors with minimal modifications. Comprehensive experiments demonstrate that our Pro3D significantly enhances existing methods, achieving state-of-the-art results on two contemporary benchmarks.

Shu Kong, Deva Ramanan

Real-world machine learning systems need to analyze test data that may differ from training data. In K-way classification, this is crisply formulated as open-set recognition, core to which is the ability to discriminate open-set data outside the K closed-set classes. Two conceptually elegant ideas for open-set discrimination are: 1) discriminatively learning an open-vs-closed binary discriminator by exploiting some outlier data as the open-set, and 2) unsupervised learning the closed-set data distribution with a GAN, using its discriminator as the open-set likelihood function. However, the former generalizes poorly to diverse open test data due to overfitting to the training outliers, which are unlikely to exhaustively span the open-world. The latter does not work well, presumably due to the instable training of GANs. Motivated by the above, we propose OpenGAN, which addresses the limitation of each approach by combining them with several technical insights. First, we show that a carefully selected GAN-discriminator on some real outlier data already achieves the state-of-the-art. Second, we augment the available set of real open training examples with adversarially synthesized "fake" data. Third and most importantly, we build the discriminator over the features computed by the closed-world K-way networks. This allows OpenGAN to be implemented via a lightweight discriminator head built on top of an existing K-way network. Extensive experiments show that OpenGAN significantly outperforms prior open-set methods.

Yi-Ting Chen, Jinghao Shi, Zelin Ye et al.

Object detection with multimodal inputs can improve many safety-critical systems such as autonomous vehicles (AVs). Motivated by AVs that operate in both day and night, we study multimodal object detection with RGB and thermal cameras, since the latter provides much stronger object signatures under poor illumination. We explore strategies for fusing information from different modalities. Our key contribution is a probabilistic ensembling technique, ProbEn, a simple non-learned method that fuses together detections from multi-modalities. We derive ProbEn from Bayes' rule and first principles that assume conditional independence across modalities. Through probabilistic marginalization, ProbEn elegantly handles missing modalities when detectors do not fire on the same object. Importantly, ProbEn also notably improves multimodal detection even when the conditional independence assumption does not hold, e.g., fusing outputs from other fusion methods (both off-the-shelf and trained in-house). We validate ProbEn on two benchmarks containing both aligned (KAIST) and unaligned (FLIR) multimodal images, showing that ProbEn outperforms prior work by more than 13% in relative performance!

Yunhan Zhao, Shu Kong, Charless Fowlkes

Monocular depth predictors are typically trained on large-scale training sets which are naturally biased w.r.t the distribution of camera poses. As a result, trained predictors fail to make reliable depth predictions for testing examples captured under uncommon camera poses. To address this issue, we propose two novel techniques that exploit the camera pose during training and prediction. First, we introduce a simple perspective-aware data augmentation that synthesizes new training examples with more diverse views by perturbing the existing ones in a geometrically consistent manner. Second, we propose a conditional model that exploits the per-image camera pose as prior knowledge by encoding it as a part of the input. We show that jointly applying the two methods improves depth prediction on images captured under uncommon and even never-before-seen camera poses. We show that our methods improve performance when applied to a range of different predictor architectures. Lastly, we show that explicitly encoding the camera pose distribution improves the generalization performance of a synthetically trained depth predictor when evaluated on real images.

Zhiyuan Fang, Shu Kong, Zhe Wang et al.

A system capturing the association between video frames and textual queries offer great potential for better video analysis. However, training such a system in a fully supervised way inevitably demands a meticulously curated video dataset with temporal-textual annotations. Therefore we provide a Weak-Supervised alternative with our proposed Referring Attention mechanism to learn temporal-textual association (dubbed WSRA). The weak supervision is simply a textual expression (e.g., short phrases or sentences) at video level, indicating this video contains relevant frames. The referring attention is our designed mechanism acting as a scoring function for grounding the given queries over frames temporally. It consists of multiple novel losses and sampling strategies for better training. The principle in our designed mechanism is to fully exploit 1) the weak supervision by considering informative and discriminative cues from intra-video segments anchored with the textual query, 2) multiple queries compared to the single video, and 3) cross-video visual similarities. We validate our WSRA through extensive experiments for temporally grounding by languages, demonstrating that it outperforms the state-of-the-art weakly-supervised methods notably.

Linfeng Wang, Shu Kong, Zachary Pincus et al.

The nematode Caenorhabditis elegans (C. elegans) serves as an important model organism in a wide variety of biological studies. In this paper we introduce a pipeline for automated analysis of C. elegans imagery for the purpose of studying life-span, health-span and the underlying genetic determinants of aging. Our system detects and segments the worm, and predicts body coordinates at each pixel location inside the worm. These coordinates provide dense correspondence across individual animals to allow for meaningful comparative analysis. We show that a model pre-trained to perform body-coordinate regression extracts rich features that can be used to predict the age of individual worms with high accuracy. This lays the ground for future research in quantifying the relation between organs' physiologic and biochemical state, and individual life/health-span.

Yunhan Zhao, Shu Kong, Daeyun Shin et al.

Leveraging synthetically rendered data offers great potential to improve monocular depth estimation and other geometric estimation tasks, but closing the synthetic-real domain gap is a non-trivial and important task. While much recent work has focused on unsupervised domain adaptation, we consider a more realistic scenario where a large amount of synthetic training data is supplemented by a small set of real images with ground-truth. In this setting, we find that existing domain translation approaches are difficult to train and offer little advantage over simple baselines that use a mix of real and synthetic data. A key failure mode is that real-world images contain novel objects and clutter not present in synthetic training. This high-level domain shift isn't handled by existing image translation models. Based on these observations, we develop an attention module that learns to identify and remove difficult out-of-domain regions in real images in order to improve depth prediction for a model trained primarily on synthetic data. We carry out extensive experiments to validate our attend-remove-complete approach (ARC) and find that it significantly outperforms state-of-the-art domain adaptation methods for depth prediction. Visualizing the removed regions provides interpretable insights into the synthetic-real domain gap.

Zhiyuan Fang, Shu Kong, Charless Fowlkes et al.

Computer Vision applications often require a textual grounding module with precision, interpretability, and resilience to counterfactual inputs/queries. To achieve high grounding precision, current textual grounding methods heavily rely on large-scale training data with manual annotations at the pixel level. Such annotations are expensive to obtain and thus severely narrow the model's scope of real-world applications. Moreover, most of these methods sacrifice interpretability, generalizability, and they neglect the importance of being resilient to counterfactual inputs. To address these issues, we propose a visual grounding system which is 1) end-to-end trainable in a weakly supervised fashion with only image-level annotations, and 2) counterfactually resilient owing to the modular design. Specifically, we decompose textual descriptions into three levels: entity, semantic attribute, color information, and perform compositional grounding progressively. We validate our model through a series of experiments and demonstrate its improvement over the state-of-the-art methods. In particular, our model's performance not only surpasses other weakly/un-supervised methods and even approaches the strongly supervised ones, but also is interpretable for decision making and performs much better in face of counterfactual classes than all the others.

Shu Kong, Charless Fowlkes

We introduce multigrid Predictive Filter Flow (mgPFF), a framework for unsupervised learning on videos. The mgPFF takes as input a pair of frames and outputs per-pixel filters to warp one frame to the other. Compared to optical flow used for warping frames, mgPFF is more powerful in modeling sub-pixel movement and dealing with corruption (e.g., motion blur). We develop a multigrid coarse-to-fine modeling strategy that avoids the requirement of learning large filters to capture large displacement. This allows us to train an extremely compact model (4.6MB) which operates in a progressive way over multiple resolutions with shared weights. We train mgPFF on unsupervised, free-form videos and show that mgPFF is able to not only estimate long-range flow for frame reconstruction and detect video shot transitions, but also readily amendable for video object segmentation and pose tracking, where it substantially outperforms the published state-of-the-art without bells and whistles. Moreover, owing to mgPFF's nature of per-pixel filter prediction, we have the unique opportunity to visualize how each pixel is evolving during solving these tasks, thus gaining better interpretability.

Shu Kong, Charless Fowlkes

We propose a simple, interpretable framework for solving a wide range of image reconstruction problems such as denoising and deconvolution. Given a corrupted input image, the model synthesizes a spatially varying linear filter which, when applied to the input image, reconstructs the desired output. The model parameters are learned using supervised or self-supervised training. We test this model on three tasks: non-uniform motion blur removal, lossy-compression artifact reduction and single image super resolution. We demonstrate that our model substantially outperforms state-of-the-art methods on all these tasks and is significantly faster than optimization-based approaches to deconvolution. Unlike models that directly predict output pixel values, the predicted filter flow is controllable and interpretable, which we demonstrate by visualizing the space of predicted filters for different tasks.

Shu Kong, Charless Fowlkes

To achieve parsimonious inference in per-pixel labeling tasks with a limited computational budget, we propose a \emph{Pixel-wise Attentional Gating} unit (\emph{PAG}) that learns to selectively process a subset of spatial locations at each layer of a deep convolutional network. PAG is a generic, architecture-independent, problem-agnostic mechanism that can be readily "plugged in" to an existing model with fine-tuning. We utilize PAG in two ways: 1) learning spatially varying pooling fields that improve model performance without the extra computation cost associated with multi-scale pooling, and 2) learning a dynamic computation policy for each pixel to decrease total computation while maintaining accuracy. We extensively evaluate PAG on a variety of per-pixel labeling tasks, including semantic segmentation, boundary detection, monocular depth and surface normal estimation. We demonstrate that PAG allows competitive or state-of-the-art performance on these tasks. Our experiments show that PAG learns dynamic spatial allocation of computation over the input image which provides better performance trade-offs compared to related approaches (e.g., truncating deep models or dynamically skipping whole layers). Generally, we observe PAG can reduce computation by $10\%$ without noticeable loss in accuracy and performance degrades gracefully when imposing stronger computational constraints.

Feng Zhou, Shu Kong, Charless Fowlkes et al.

Automated facial expression analysis has a variety of applications in human-computer interaction. Traditional methods mainly analyze prototypical facial expressions of no more than eight discrete emotions as a classification task. However, in practice, spontaneous facial expressions in naturalistic environment can represent not only a wide range of emotions, but also different intensities within an emotion family. In such situation, these methods are not reliable or adequate. In this paper, we propose to train deep convolutional neural networks (CNNs) to analyze facial expressions explainable in a dimensional emotion model. The proposed method accommodates not only a set of basic emotion expressions, but also a full range of other emotions and subtle emotion intensities that we both feel in ourselves and perceive in others in our daily life. Specifically, we first mapped facial expressions into dimensional measures so that we transformed facial expression analysis from a classification problem to a regression one. We then tested our CNN-based methods for facial expression regression and these methods demonstrated promising performance. Moreover, we improved our method by a bilinear pooling which encodes second-order statistics of features. We showed such bilinear-CNN models significantly outperformed their respective baselines.

Zhiyuan Fang, Shu Kong, Tianshu Yu et al.

Grounding textual phrases in visual content is a meaningful yet challenging problem with various potential applications such as image-text inference or text-driven multimedia interaction. Most of the current existing methods adopt the supervised learning mechanism which requires ground-truth at pixel level during training. However, fine-grained level ground-truth annotation is quite time-consuming and severely narrows the scope for more general applications. In this extended abstract, we explore methods to localize flexibly image regions from the top-down signal (in a form of one-hot label or natural languages) with a weakly supervised attention learning mechanism. In our model, two types of modules are utilized: a backbone module for visual feature capturing, and an attentive module generating maps based on regularized bilinear pooling. We construct the model in an end-to-end fashion which is trained by encouraging the spatial attentive map to shift and focus on the region that consists of the best matched visual features with the top-down signal. We demonstrate the preliminary yet promising results on a testbed that is synthesized with multi-label MNIST data.

Shu Kong, Charless Fowlkes

We introduce a differentiable, end-to-end trainable framework for solving pixel-level grouping problems such as instance segmentation consisting of two novel components. First, we regress pixels into a hyper-spherical embedding space so that pixels from the same group have high cosine similarity while those from different groups have similarity below a specified margin. We analyze the choice of embedding dimension and margin, relating them to theoretical results on the problem of distributing points uniformly on the sphere. Second, to group instances, we utilize a variant of mean-shift clustering, implemented as a recurrent neural network parameterized by kernel bandwidth. This recurrent grouping module is differentiable, enjoys convergent dynamics and probabilistic interpretability. Backpropagating the group-weighted loss through this module allows learning to focus on only correcting embedding errors that won't be resolved during subsequent clustering. Our framework, while conceptually simple and theoretically abundant, is also practically effective and computationally efficient. We demonstrate substantial improvements over state-of-the-art instance segmentation for object proposal generation, as well as demonstrating the benefits of grouping loss for classification tasks such as boundary detection and semantic segmentation.

Shu Kong, Charless Fowlkes

Objects may appear at arbitrary scales in perspective images of a scene, posing a challenge for recognition systems that process images at a fixed resolution. We propose a depth-aware gating module that adaptively selects the pooling field size in a convolutional network architecture according to the object scale (inversely proportional to the depth) so that small details are preserved for distant objects while larger receptive fields are used for those nearby. The depth gating signal is provided by stereo disparity or estimated directly from monocular input. We integrate this depth-aware gating into a recurrent convolutional neural network to perform semantic segmentation. Our recurrent module iteratively refines the segmentation results, leveraging the depth and semantic predictions from the previous iterations. Through extensive experiments on four popular large-scale RGB-D datasets, we demonstrate this approach achieves competitive semantic segmentation performance with a model which is substantially more compact. We carry out extensive analysis of this architecture including variants that operate on monocular RGB but use depth as side-information during training, unsupervised gating as a generic attentional mechanism, and multi-resolution gating. We find that gated pooling for joint semantic segmentation and depth yields state-of-the-art results for quantitative monocular depth estimation.

Shu Kong, Charless Fowlkes

Pooling second-order local feature statistics to form a high-dimensional bilinear feature has been shown to achieve state-of-the-art performance on a variety of fine-grained classification tasks. To address the computational demands of high feature dimensionality, we propose to represent the covariance features as a matrix and apply a low-rank bilinear classifier. The resulting classifier can be evaluated without explicitly computing the bilinear feature map which allows for a large reduction in the compute time as well as decreasing the effective number of parameters to be learned. To further compress the model, we propose classifier co-decomposition that factorizes the collection of bilinear classifiers into a common factor and compact per-class terms. The co-decomposition idea can be deployed through two convolutional layers and trained in an end-to-end architecture. We suggest a simple yet effective initialization that avoids explicitly first training and factorizing the larger bilinear classifiers. Through extensive experiments, we show that our model achieves state-of-the-art performance on several public datasets for fine-grained classification trained with only category labels. Importantly, our final model is an order of magnitude smaller than the recently proposed compact bilinear model, and three orders smaller than the standard bilinear CNN model.

Shu Kong, Xiaohui Shen, Zhe Lin et al.

Real-world applications could benefit from the ability to automatically generate a fine-grained ranking of photo aesthetics. However, previous methods for image aesthetics analysis have primarily focused on the coarse, binary categorization of images into high- or low-aesthetic categories. In this work, we propose to learn a deep convolutional neural network to rank photo aesthetics in which the relative ranking of photo aesthetics are directly modeled in the loss function. Our model incorporates joint learning of meaningful photographic attributes and image content information which can help regularize the complicated photo aesthetics rating problem. To train and analyze this model, we have assembled a new aesthetics and attributes database (AADB) which contains aesthetic scores and meaningful attributes assigned to each image by multiple human raters. Anonymized rater identities are recorded across images allowing us to exploit intra-rater consistency using a novel sampling strategy when computing the ranking loss of training image pairs. We show the proposed sampling strategy is very effective and robust in face of subjective judgement of image aesthetics by individuals with different aesthetic tastes. Experiments demonstrate that our unified model can generate aesthetic rankings that are more consistent with human ratings. To further validate our model, we show that by simply thresholding the estimated aesthetic scores, we are able to achieve state-or-the-art classification performance on the existing AVA dataset benchmark.

Shu Kong, Surangi Punyasena, Charless Fowlkes

We propose a robust approach for performing automatic species-level recognition of fossil pollen grains in microscopy images that exploits both global shape and local texture characteristics in a patch-based matching methodology. We introduce a novel criteria for selecting meaningful and discriminative exemplar patches. We optimize this function during training using a greedy submodular function optimization framework that gives a near-optimal solution with bounded approximation error. We use these selected exemplars as a dictionary basis and propose a spatially-aware sparse coding method to match testing images for identification while maintaining global shape correspondence. To accelerate the coding process for fast matching, we introduce a relaxed form that uses spatially-aware soft-thresholding during coding. Finally, we carry out an experimental study that demonstrates the effectiveness and efficiency of our exemplar selection and classification mechanisms, achieving $86.13\%$ accuracy on a difficult fine-grained species classification task distinguishing three types of fossil spruce pollen.

Shu Kong, Zhuolin Jiang, Qiang Yang

The challenge of object categorization in images is largely due to arbitrary translations and scales of the foreground objects. To attack this difficulty, we propose a new approach called collaborative receptive field learning to extract specific receptive fields (RF's) or regions from multiple images, and the selected RF's are supposed to focus on the foreground objects of a common category. To this end, we solve the problem by maximizing a submodular function over a similarity graph constructed by a pool of RF candidates. However, measuring pairwise distance of RF's for building the similarity graph is a nontrivial problem. Hence, we introduce a similarity metric called pyramid-error distance (PED) to measure their pairwise distances through summing up pyramid-like matching errors over a set of low-level features. Besides, in consistent with the proposed PED, we construct a simple nonparametric classifier for classification. Experimental results show that our method effectively discovers the foreground objects in images, and improves classification performance.

Shu Kong, Zhuolin Jiang, Qiang Yang

We now know that mid-level features can greatly enhance the performance of image learning, but how to automatically learn the image features efficiently and in an unsupervised manner is still an open question. In this paper, we present a very efficient mid-level feature learning approach (MidFea), which only involves simple operations such as $k$-means clustering, convolution, pooling, vector quantization and random projection. We explain why this simple method generates the desired features, and argue that there is no need to spend much time in learning low-level feature extractors. Furthermore, to boost the performance, we propose to model the neuron selectivity (NS) principle by building an additional layer over the mid-level features before feeding the features into the classifier. We show that the NS-layer learns category-specific neurons with both bottom-up inference and top-down analysis, and thus supports fast inference for a query image. We run extensive experiments on several public databases to demonstrate that our approach can achieve state-of-the-art performances for face recognition, gender classification, age estimation and object categorization. In particular, we demonstrate that our approach is more than an order of magnitude faster than some recently proposed sparse coding based methods.