Hang Zhang

Hang Zhang, Xin Li, Lidong Bing

We present Video-LLaMA a multi-modal framework that empowers Large Language Models (LLMs) with the capability of understanding both visual and auditory content in the video. Video-LLaMA bootstraps cross-modal training from the frozen pre-trained visual and audio encoders and the frozen LLMs. Unlike previous works that complement LLMs to process the visual or audio signals only, Video-LLaMA enables video comprehension by tackling two challenges: (1) capturing the temporal changes in visual scenes, (2) integrating audio-visual signals. To counter the first challenge, we propose a Video Q-former to assemble a pre-trained image encoder into our video encoder and introduce a video-to-text generation task to learn video-language correspondence. For the second challenge, we leverage ImageBind, a universal embedding model aligning multiple modalities, as the pre-trained audio encoder and introduce an Audio Q-former on top of ImageBind to learn reasonable auditory query embeddings for the LLM module. To align the output of both visual and audio encoders with LLM's embedding space, we first train Video-LLaMA on massive video/image-caption pairs and then tune our model with visual-instruction datasets of moderate amount but higher quality. We found Video-LLaMA shows the ability to perceive and comprehend video content and generate meaningful responses grounded in the visual and auditory information presented in the videos.

Feng Liang, Bichen Wu, Xiaoliang Dai et al.

Open-vocabulary semantic segmentation aims to segment an image into semantic regions according to text descriptions, which may not have been seen during training. Recent two-stage methods first generate class-agnostic mask proposals and then leverage pre-trained vision-language models, e.g., CLIP, to classify masked regions. We identify the performance bottleneck of this paradigm to be the pre-trained CLIP model, since it does not perform well on masked images. To address this, we propose to finetune CLIP on a collection of masked image regions and their corresponding text descriptions. We collect training data by mining an existing image-caption dataset (e.g., COCO Captions), using CLIP to match masked image regions to nouns in the image captions. Compared with the more precise and manually annotated segmentation labels with fixed classes (e.g., COCO-Stuff), we find our noisy but diverse dataset can better retain CLIP's generalization ability. Along with finetuning the entire model, we utilize the "blank" areas in masked images using a method we dub mask prompt tuning. Experiments demonstrate mask prompt tuning brings significant improvement without modifying any weights of CLIP, and it can further improve a fully finetuned model. In particular, when trained on COCO and evaluated on ADE20K-150, our best model achieves 29.6% mIoU, which is +8.5% higher than the previous state-of-the-art. For the first time, open-vocabulary generalist models match the performance of supervised specialist models in 2017 without dataset-specific adaptations.

Hang Zhang, Renjiu Hu, Xiang Chen et al.

Recent research highlights that the Directed Accumulator (DA), through its parametrization of geometric priors into neural networks, has notably improved the performance of medical image recognition, particularly with small and imbalanced datasets. However, DA's potential in pixel-wise dense predictions is unexplored. To bridge this gap, we present the Directed Accumulator Grid (DAGrid), which allows geometric-preserving filtering in neural networks, thus broadening the scope of DA's applications to include pixel-level dense prediction tasks. DAGrid utilizes homogeneous data types in conjunction with designed sampling grids to construct geometrically transformed representations, retaining intricate geometric information and promoting long-range information propagation within the neural networks. Contrary to its symmetric counterpart, grid sampling, which might lose information in the sampling process, DAGrid aggregates all pixels, ensuring a comprehensive representation in the transformed space. The parallelization of DAGrid on modern GPUs is facilitated using CUDA programming, and also back propagation is enabled for deep neural network training. Empirical results show DAGrid-enhanced neural networks excel in supervised skin lesion segmentation and unsupervised cardiac image registration. Specifically, the network incorporating DAGrid has realized a 70.8% reduction in network parameter size and a 96.8% decrease in FLOPs, while concurrently improving the Dice score for skin lesion segmentation by 1.0% compared to state-of-the-art transformers. Furthermore, it has achieved improvements of 4.4% and 8.2% in the average Dice score and Dice score of the left ventricular mass, respectively, indicating an increase in registration accuracy for cardiac images. The source code is available at https://github.com/tinymilky/DeDA.

Xingwei He, Yeyun Gong, A-Long Jin et al.

The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query.In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality query-informed document representations. Experimental results on both in-domain and out-of-domain datasets demonstrate that our approach outperforms previous dense retrieval models.

Hang Zhang

In this research, I proposed a network structure for multi-view 3D object detection using camera-only data and a Bird's-Eye-View map. My work is based on a current key challenge domain adaptation and visual data transfer. Although many excellent camera-only 3D object detection has been continuously proposed, many research work risk dramatic performance drop when the networks are trained on the source domain but tested on a different target domain. Then I found it is very surprising that predictions on bounding boxes and classes are still replied to on 2D networks. Based on the domain gap assumption on various 3D datasets, I found they still shared a similar data extraction on the same BEV map size and camera data transfer. Therefore, to analyze the domain gap influence on the current method and to make good use of 3D space information among the dataset and the real world, I proposed a transfer learning method and Transformer construction to study the 3D object detection on NuScenes-mini and Lyft. Through multi-dataset training and a detection head from the Transformer, the network demonstrated good data migration performance and efficient detection performance by using 3D anchor query and 3D positional information. Relying on only a small amount of source data and the existing large model pre-training weights, the efficient network manages to achieve competitive results on the new target domain. Moreover, my study utilizes 3D information as available semantic information and 2D multi-view image features blending into the visual-language transfer design. In the final 3D anchor box prediction and object classification, my network achieved good results on standard metrics of 3D object detection, which differs from dataset-specific models on each training domain without any fine-tuning.

Hang Zhang, Rongguang Wang, Jinwei Zhang et al.

We introduce Neural Representation of Distribution (NeRD) technique, a module for convolutional neural networks (CNNs) that can estimate the feature distribution by optimizing an underlying function mapping image coordinates to the feature distribution. Using NeRD, we propose an end-to-end deep learning model for medical image segmentation that can compensate the negative impact of feature distribution shifting issue caused by commonly used network operations such as padding and pooling. An implicit function is used to represent the parameter space of the feature distribution by querying the image coordinate. With NeRD, the impact of issues such as over-segmenting and missing have been reduced, and experimental results on the challenging white matter lesion segmentation and left atrial segmentation verify the effectiveness of the proposed method. The code is available via https://github.com/tinymilky/NeRD.

Mingqi Shao, Feng Xiong, Hang Zhang et al.

Neural radiance fields~(NeRF) have recently been applied to render large-scale scenes. However, their limited model capacity typically results in blurred rendering results. Existing large-scale NeRFs primarily address this limitation by partitioning the scene into blocks, which are subsequently handled by separate sub-NeRFs. These sub-NeRFs, trained from scratch and processed independently, lead to inconsistencies in geometry and appearance across the scene. Consequently, the rendering quality fails to exhibit significant improvement despite the expansion of model capacity. In this work, we present global-guided focal neural radiance field (GF-NeRF) that achieves high-fidelity rendering of large-scale scenes. Our proposed GF-NeRF utilizes a two-stage (Global and Focal) architecture and a global-guided training strategy. The global stage obtains a continuous representation of the entire scene while the focal stage decomposes the scene into multiple blocks and further processes them with distinct sub-encoders. Leveraging this two-stage architecture, sub-encoders only need fine-tuning based on the global encoder, thus reducing training complexity in the focal stage while maintaining scene-wide consistency. Spatial information and error information from the global stage also benefit the sub-encoders to focus on crucial areas and effectively capture more details of large-scale scenes. Notably, our approach does not rely on any prior knowledge about the target scene, attributing GF-NeRF adaptable to various large-scale scene types, including street-view and aerial-view scenes. We demonstrate that our method achieves high-fidelity, natural rendering results on various types of large-scale datasets. Our project page: https://shaomq2187.github.io/GF-NeRF/

Bichen Wu, Chaojian Li, Hang Zhang et al.

Neural Architecture Search (NAS) has been widely adopted to design accurate and efficient image classification models. However, applying NAS to a new computer vision task still requires a huge amount of effort. This is because 1) previous NAS research has been over-prioritized on image classification while largely ignoring other tasks; 2) many NAS works focus on optimizing task-specific components that cannot be favorably transferred to other tasks; and 3) existing NAS methods are typically designed to be "proxyless" and require significant effort to be integrated with each new task's training pipelines. To tackle these challenges, we propose FBNetV5, a NAS framework that can search for neural architectures for a variety of vision tasks with much reduced computational cost and human effort. Specifically, we design 1) a search space that is simple yet inclusive and transferable; 2) a multitask search process that is disentangled with target tasks' training pipeline; and 3) an algorithm to simultaneously search for architectures for multiple tasks with a computational cost agnostic to the number of tasks. We evaluate the proposed FBNetV5 targeting three fundamental vision tasks -- image classification, object detection, and semantic segmentation. Models searched by FBNetV5 in a single run of search have outperformed the previous stateof-the-art in all the three tasks: image classification (e.g., +1.3% ImageNet top-1 accuracy under the same FLOPs as compared to FBNetV3), semantic segmentation (e.g., +1.8% higher ADE20K val. mIoU than SegFormer with 3.6x fewer FLOPs), and object detection (e.g., +1.1% COCO val. mAP with 1.2x fewer FLOPs as compared to YOLOX).

Jia Xue, Hang Zhang, Ko Nishino et al.

Computational surface modeling that underlies material recognition has transitioned from reflectance modeling using in-lab controlled radiometric measurements to image-based representations based on internet-mined single-view images captured in the scene. We take a middle-ground approach for material recognition that takes advantage of both rich radiometric cues and flexible image capture. A key concept is differential angular imaging, where small angular variations in image capture enables angular-gradient features for an enhanced appearance representation that improves recognition. We build a large-scale material database, Ground Terrain in Outdoor Scenes (GTOS) database, to support ground terrain recognition for applications such as autonomous driving and robot navigation. The database consists of over 30,000 images covering 40 classes of outdoor ground terrain under varying weather and lighting conditions. We develop a novel approach for material recognition called texture-encoded angular network (TEAN) that combines deep encoding pooling of RGB information and differential angular images for angular-gradient features to fully leverage this large dataset. With this novel network architecture, we extract characteristics of materials encoded in the angular and spatial gradients of their appearance. Our results show that TEAN achieves recognition performance that surpasses single view performance and standard (non-differential/large-angle sampling) multiview performance.

Shijie Geng, Ji Zhang, Zuohui Fu et al.

Different from short videos and GIFs, video stories contain clear plots and lists of principal characters. Without identifying the connection between appearing people and character names, a model is not able to obtain a genuine understanding of the plots. Video Story Question Answering (VSQA) offers an effective way to benchmark higher-level comprehension abilities of a model. However, current VSQA methods merely extract generic visual features from a scene. With such an approach, they remain prone to learning just superficial correlations. In order to attain a genuine understanding of who did what to whom, we propose a novel model that continuously refines character-aware relations. This model specifically considers the characters in a video story, as well as the relations connecting different characters and objects. Based on these signals, our framework enables weakly-supervised face naming through multi-instance co-occurrence matching and supports high-level reasoning utilizing Transformer structures. We train and test our model on the six diverse TV shows in the TVQA dataset, which is by far the largest and only publicly available dataset for VSQA. We validate our proposed approach over TVQA dataset through extensive ablation study.

Yi Zhu, Zhongyue Zhang, Chongruo Wu et al.

Deep learning usually achieves the best results with complete supervision. In the case of semantic segmentation, this means that large amounts of pixelwise annotations are required to learn accurate models. In this paper, we show that we can obtain state-of-the-art results using a semi-supervised approach, specifically a self-training paradigm. We first train a teacher model on labeled data, and then generate pseudo labels on a large set of unlabeled data. Our robust training framework can digest human-annotated and pseudo labels jointly and achieve top performances on Cityscapes, CamVid and KITTI datasets while requiring significantly less supervision. We also demonstrate the effectiveness of self-training on a challenging cross-domain generalization task, outperforming conventional finetuning method by a large margin. Lastly, to alleviate the computational burden caused by the large amount of pseudo labels, we propose a fast training schedule to accelerate the training of segmentation models by up to 2x without performance degradation.

Shijie Geng, Ji Zhang, Hang Zhang et al.

We present a simple method that achieves unexpectedly superior performance for Complex Reasoning involved Visual Question Answering. Our solution collects statistical features from high-frequency words of all the questions asked about an image and use them as accurate knowledge for answering further questions of the same image. We are fully aware that this setting is not ubiquitously applicable, and in a more common setting one should assume the questions are asked separately and they cannot be gathered to obtain a knowledge base. Nonetheless, we use this method as an evidence to demonstrate our observation that the bottleneck effect is more severe on the feature extraction part than it is on the knowledge reasoning part. We show significant gaps when using the same reasoning model with 1) ground-truth features; 2) statistical features; 3) detected features from completely learned detectors, and analyze what these gaps mean to researches on visual reasoning topics. Our model with the statistical features achieves the 2nd place in the GQA Challenge 2019.

Jian Guo, He He, Tong He et al.

We present GluonCV and GluonNLP, the deep learning toolkits for computer vision and natural language processing based on Apache MXNet (incubating). These toolkits provide state-of-the-art pre-trained models, training scripts, and training logs, to facilitate rapid prototyping and promote reproducible research. We also provide modular APIs with flexible building blocks to enable efficient customization. Leveraging the MXNet ecosystem, the deep learning models in GluonCV and GluonNLP can be deployed onto a variety of platforms with different programming languages. The Apache 2.0 license has been adopted by GluonCV and GluonNLP to allow for software distribution, modification, and usage.

Haibin Lin, Hang Zhang, Yifei Ma et al.

With an increasing demand for training powers for deep learning algorithms and the rapid growth of computation resources in data centers, it is desirable to dynamically schedule different distributed deep learning tasks to maximize resource utilization and reduce cost. In this process, different tasks may receive varying numbers of machines at different time, a setting we call elastic distributed training. Despite the recent successes in large mini-batch distributed training, these methods are rarely tested in elastic distributed training environments and suffer degraded performance in our experiments, when we adjust the learning rate linearly immediately with respect to the batch size. One difficulty we observe is that the noise in the stochastic momentum estimation is accumulated over time and will have delayed effects when the batch size changes. We therefore propose to smoothly adjust the learning rate over time to alleviate the influence of the noisy momentum estimation. Our experiments on image classification, object detection and semantic segmentation have demonstrated that our proposed Dynamic SGD method achieves stabilized performance when varying the number of GPUs from 8 to 128. We also provide theoretical understanding on the optimality of linear learning rate scheduling and the effects of stochastic momentum.

Zhi Zhang, Tong He, Hang Zhang et al.

Training heuristics greatly improve various image classification model accuracies~\cite{he2018bag}. Object detection models, however, have more complex neural network structures and optimization targets. The training strategies and pipelines dramatically vary among different models. In this works, we explore training tweaks that apply to various models including Faster R-CNN and YOLOv3. These tweaks do not change the model architectures, therefore, the inference costs remain the same. Our empirical results demonstrate that, however, these freebies can improve up to 5% absolute precision compared to state-of-the-art baselines.

Tong He, Zhi Zhang, Hang Zhang et al.

Much of the recent progress made in image classification research can be credited to training procedure refinements, such as changes in data augmentations and optimization methods. In the literature, however, most refinements are either briefly mentioned as implementation details or only visible in source code. In this paper, we will examine a collection of such refinements and empirically evaluate their impact on the final model accuracy through ablation study. We will show that, by combining these refinements together, we are able to improve various CNN models significantly. For example, we raise ResNet-50's top-1 validation accuracy from 75.3% to 79.29% on ImageNet. We will also demonstrate that improvement on image classification accuracy leads to better transfer learning performance in other application domains such as object detection and semantic segmentation.

Jia Xue, Hang Zhang, Kristin Dana

We present a texture network called Deep Encoding Pooling Network (DEP) for the task of ground terrain recognition. Recognition of ground terrain is an important task in establishing robot or vehicular control parameters, as well as for localization within an outdoor environment. The architecture of DEP integrates orderless texture details and local spatial information and the performance of DEP surpasses state-of-the-art methods for this task. The GTOS database (comprised of over 30,000 images of 40 classes of ground terrain in outdoor scenes) enables supervised recognition. For evaluation under realistic conditions, we use test images that are not from the existing GTOS dataset, but are instead from hand-held mobile phone videos of similar terrain. This new evaluation dataset, GTOS-mobile, consists of 81 videos of 31 classes of ground terrain such as grass, gravel, asphalt and sand. The resultant network shows excellent performance not only for GTOS-mobile, but also for more general databases (MINC and DTD). Leveraging the discriminant features learned from this network, we build a new texture manifold called DEP-manifold. We learn a parametric distribution in feature space in a fully supervised manner, which gives the distance relationship among classes and provides a means to implicitly represent ambiguous class boundaries. The source code and database are publicly available.

Hang Zhang, Kristin Dana, Jianping Shi et al.

Recent work has made significant progress in improving spatial resolution for pixelwise labeling with Fully Convolutional Network (FCN) framework by employing Dilated/Atrous convolution, utilizing multi-scale features and refining boundaries. In this paper, we explore the impact of global contextual information in semantic segmentation by introducing the Context Encoding Module, which captures the semantic context of scenes and selectively highlights class-dependent featuremaps. The proposed Context Encoding Module significantly improves semantic segmentation results with only marginal extra computation cost over FCN. Our approach has achieved new state-of-the-art results 51.7% mIoU on PASCAL-Context, 85.9% mIoU on PASCAL VOC 2012. Our single model achieves a final score of 0.5567 on ADE20K test set, which surpass the winning entry of COCO-Place Challenge in 2017. In addition, we also explore how the Context Encoding Module can improve the feature representation of relatively shallow networks for the image classification on CIFAR-10 dataset. Our 14 layer network has achieved an error rate of 3.45%, which is comparable with state-of-the-art approaches with over 10 times more layers. The source code for the complete system are publicly available.

Hang Zhang, Kristin Dana

Despite the rapid progress in style transfer, existing approaches using feed-forward generative network for multi-style or arbitrary-style transfer are usually compromised of image quality and model flexibility. We find it is fundamentally difficult to achieve comprehensive style modeling using 1-dimensional style embedding. Motivated by this, we introduce CoMatch Layer that learns to match the second order feature statistics with the target styles. With the CoMatch Layer, we build a Multi-style Generative Network (MSG-Net), which achieves real-time performance. We also employ an specific strategy of upsampled convolution which avoids checkerboard artifacts caused by fractionally-strided convolution. Our method has achieved superior image quality comparing to state-of-the-art approaches. The proposed MSG-Net as a general approach for real-time style transfer is compatible with most existing techniques including content-style interpolation, color-preserving, spatial control and brush stroke size control. MSG-Net is the first to achieve real-time brush-size control in a purely feed-forward manner for style transfer. Our implementations and pre-trained models for Torch, PyTorch and MXNet frameworks will be publicly available.

Hang Zhang, Jia Xue, Kristin Dana

We propose a Deep Texture Encoding Network (Deep-TEN) with a novel Encoding Layer integrated on top of convolutional layers, which ports the entire dictionary learning and encoding pipeline into a single model. Current methods build from distinct components, using standard encoders with separate off-the-shelf features such as SIFT descriptors or pre-trained CNN features for material recognition. Our new approach provides an end-to-end learning framework, where the inherent visual vocabularies are learned directly from the loss function. The features, dictionaries and the encoding representation for the classifier are all learned simultaneously. The representation is orderless and therefore is particularly useful for material and texture recognition. The Encoding Layer generalizes robust residual encoders such as VLAD and Fisher Vectors, and has the property of discarding domain specific information which makes the learned convolutional features easier to transfer. Additionally, joint training using multiple datasets of varied sizes and class labels is supported resulting in increased recognition performance. The experimental results show superior performance as compared to state-of-the-art methods using gold-standard databases such as MINC-2500, Flickr Material Database, KTH-TIPS-2b, and two recent databases 4D-Light-Field-Material and GTOS. The source code for the complete system are publicly available.

Jia Xue, Hang Zhang, Kristin Dana et al.

Material recognition for real-world outdoor surfaces has become increasingly important for computer vision to support its operation "in the wild." Computational surface modeling that underlies material recognition has transitioned from reflectance modeling using in-lab controlled radiometric measurements to image-based representations based on internet-mined images of materials captured in the scene. We propose to take a middle-ground approach for material recognition that takes advantage of both rich radiometric cues and flexible image capture. We realize this by developing a framework for differential angular imaging, where small angular variations in image capture provide an enhanced appearance representation and significant recognition improvement. We build a large-scale material database, Ground Terrain in Outdoor Scenes (GTOS) database, geared towards real use for autonomous agents. The database consists of over 30,000 images covering 40 classes of outdoor ground terrain under varying weather and lighting conditions. We develop a novel approach for material recognition called a Differential Angular Imaging Network (DAIN) to fully leverage this large dataset. With this novel network architecture, we extract characteristics of materials encoded in the angular and spatial gradients of their appearance. Our results show that DAIN achieves recognition performance that surpasses single view or coarsely quantized multiview images. These results demonstrate the effectiveness of differential angular imaging as a means for flexible, in-place material recognition.

Hang Zhang, Kristin Dana, Ko Nishino

Images are the standard input for vision algorithms, but one-shot infield reflectance measurements are creating new opportunities for recognition and scene understanding. In this work, we address the question of what reflectance can reveal about materials in an efficient manner. We go beyond the question of recognition and labeling and ask the question: What intrinsic physical properties of the surface can be estimated using reflectance? We introduce a framework that enables prediction of actual friction values for surfaces using one-shot reflectance measurements. This work is a first of its kind vision-based friction estimation. We develop a novel representation for reflectance disks that capture partial BRDF measurements instantaneously. Our method of deep reflectance codes combines CNN features and fisher vector pooling with optimal binary embedding to create codes that have sufficient discriminatory power and have important properties of illumination and spatial invariance. The experimental results demonstrate that reflectance can play a new role in deciphering the underlying physical properties of real-world scenes.