Long Mai

Long Mai, Julie Carson-Berndsen

The integration of natural language processing (NLP) technologies into educational applications has shown promising results, particularly in the language learning domain. Recently, many spoken open-domain chatbots have been used as speaking partners, helping language learners improve their language skills. However, one of the significant challenges is the high word-error-rate (WER) when recognizing non-native/non-fluent speech, which interrupts conversation flow and leads to disappointment for learners. This paper explores the use of GPT4 for ASR error correction in conversational settings. In addition to WER, we propose to use semantic textual similarity (STS) and next response sensibility (NRS) metrics to evaluate the impact of error correction models on the quality of the conversation. We find that transcriptions corrected by GPT4 lead to higher conversation quality, despite an increase in WER. GPT4 also outperforms standard error correction methods without the need for in-domain training data.

Long Mai, Julie Carson-Berndsen

The transcription quality of automatic speech recognition (ASR) systems degrades significantly when transcribing audios coming from unseen domains. We propose an unsupervised error correction method for unsupervised ASR domain adaption, aiming to recover transcription errors caused by domain mismatch. Unlike existing correction methods that rely on transcribed audios for training, our approach requires only unlabeled data of the target domains in which a pseudo-labeling technique is applied to generate correction training samples. To reduce over-fitting to the pseudo data, we also propose an encoder-decoder correction model that can take into account additional information such as dialogue context and acoustic features. Experiment results show that our method obtains a significant word error rate (WER) reduction over non-adapted ASR systems. The correction model can also be applied on top of other adaptation approaches to bring an additional improvement of 10% relatively.

Yan-Bo Lin, Jonah Casebeer, Long Mai et al.

Generating music that temporally aligns with video events is challenging for existing text-to-music models, which lack fine-grained temporal control. We introduce V2M-Zero, a zero-pair video-to-music generation approach that outputs time-aligned music for video. Our method is motivated by a key observation: temporal synchronization requires matching when and how much change occurs, not what changes. While musical and visual events differ semantically, they exhibit shared temporal structure that can be captured independently within each modality. We capture this structure through event curves computed from intra-modal similarity using pretrained music and video encoders. By measuring temporal change within each modality independently, these curves provide comparable representations across modalities. This enables a simple training strategy: fine-tune a text-to-music model on music-event curves, then substitute video-event curves at inference without cross-modal training or paired data. Across OES-Pub, MovieGenBench-Music, and AIST++, V2M-Zero achieves substantial gains over paired-data baselines: 5-21% higher audio quality, 13-15% better semantic alignment, 21-52% improved temporal synchronization, and 28% higher beat alignment on dance videos. We find similar results via a large crowd-source subjective listening test. Overall, our results validate that temporal alignment through within-modality features, rather than paired cross-modal supervision, is effective for video-to-music generation. Results are available at https://genjib.github.io/v2m_zero/

Long Mai

Real-time spoken dialogue systems face a fundamental tension between latency and response quality. End-to-end speech-to-speech (S2S) models respond immediately and naturally handle turn-taking, backchanneling, and interruption, but produce semantically weaker outputs. Cascaded pipelines (ASR -> LLM) deliver stronger responses at the cost of latency that grows with model size. We present RelayS2S, a hybrid architecture that runs two paths in parallel upon turn detection. The fast path -- a duplex S2S model -- speculatively drafts a short response prefix that is streamed immediately to TTS for low-latency audio onset, while continuing to monitor live audio events. The slow path -- a cascaded ASR -> LLM pipeline -- generates a higher-quality continuation conditioned on the committed prefix, producing a seamless utterance. A lightweight learned verifier gates the handoff, committing the prefix when appropriate or falling back gracefully to the slow path alone. Experiments show that RelayS2S achieves P90 onset latency comparable to the S2S model while retaining 99% cascaded response quality in average score, with benefits growing as the slow-path model scales. Because the prefix handoff requires no architectural modification to either component, RelayS2S serves as a lightweight, drop-in addition to existing cascaded pipelines. Our code and data are publicly available at: https://github.com/mailong25/relays2s

Quynh Phung, Long Mai, Cusuh Ham et al.

We study object motion path editing in videos, where the goal is to alter a target object's trajectory while preserving the original scene content. Unlike prior video editing methods that primarily manipulate appearance or rely on point-track-based trajectory control, which is often challenging for users to provide during inference, especially in videos with camera motion, we offer a practical, easy-to-use approach to controllable object-centric motion editing. We present Trace, a framework that enables users to design the desired trajectory in a single anchor frame and then synthesizes a temporally consistent edited video. Our approach addresses this task with a two-stage pipeline: a cross-view motion transformation module that maps first-frame path design to frame-aligned box trajectories under camera motion, and a motion-conditioned video re-synthesis module that follows these trajectories to regenerate the object while preserving the remaining content of the input video. Experiments on diverse real-world videos show that our method produces more coherent, realistic, and controllable motion edits than recent image-to-video and video-to-video methods.

Elena Shushkevich, Long Mai, Manuel V. Loureiro et al.

The proliferation of news media outlets has increased the demand for intelligent systems capable of detecting redundant information in news articles in order to enhance user experience. However, the heterogeneous nature of news can lead to spurious findings in these systems: Simple heuristics such as whether a pair of news are both about politics can provide strong but deceptive downstream performance. Segmenting news similarity datasets into topics improves the training of these models by forcing them to learn how to distinguish salient characteristics under more narrow domains. However, this requires the existence of topic-specific datasets, which are currently lacking. In this article, we propose a novel dataset of similar news, SPICED, which includes seven topics: Crime & Law, Culture & Entertainment, Disasters & Accidents, Economy & Business, Politics & Conflicts, Science & Technology, and Sports. Futhermore, we present four different levels of complexity, specifically designed for news similarity detection task. We benchmarked the created datasets using MinHash, BERT, SBERT, and SimCSE models.

Long Mai, Julie Carson-Berndsen

Recent advancements in large language models (LLMs) have led to significant progress in text-based dialogue systems. These systems can now generate high-quality responses that are accurate and coherent across a wide range of topics and tasks. However, spoken dialogue systems still lag behind in terms of naturalness. They tend to produce robotic interactions, with issues such as slow response times, overly generic or cautious replies, and a lack of natural rhythm and fluid turn-taking. This shortcoming is largely due to the over-reliance on the traditional cascaded design, which involve separate, sequential components, as well as the use of text as an intermediate representation. This paper propose a real-time, textless spoken dialogue generation model (RTTL-DG) that aims to overcome these challenges. Our system enables fluid turn-taking and generates responses with minimal delay by processing streaming spoken conversation directly. Additionally, our model incorporates backchannels, filters, laughter, and other paralinguistic signals, which are often absent in cascaded dialogue systems, to create more natural and human-like interactions. The implementations and generated samples are available in our repository: https://github.com/mailong25/rts2s-dg

Zhixuan Liu, Peter Schaldenbrand, Yijun Li et al.

We present TokenDial, a framework for continuous, slider-style attribute control in pretrained text-to-video generation models. While modern generators produce strong holistic videos, they offer limited control over how much an attribute changes (e.g., effect intensity or motion magnitude) without drifting identity, background, or temporal coherence. TokenDial is built on the observation: additive offsets in the intermediate spatiotemporal visual patch-token space form a semantic control direction, where adjusting the offset magnitude yields coherent, predictable edits for both appearance and motion dynamics. We learn attribute-specific token offsets without retraining the backbone, using pretrained understanding signals: semantic direction matching for appearance and motion-magnitude scaling for motion. We demonstrate TokenDial's effectiveness on diverse attributes and prompts, achieving stronger controllability and higher-quality edits than state-of-the-art baselines, supported by extensive quantitative evaluation and human studies.

Long Mai, Julie Carson-Berndsen

While Large Language Models (LLMs) have made significant strides in replicating human-like abilities, there are concerns about a reduction in the linguistic diversity of their outputs. This results in the homogenization of viewpoints and perspectives, as well as the underrepresentation of specific demographic groups. Although several fine-tuning and prompting techniques have been suggested to tackle the issue, they are often tailored to specific tasks or come with a substantial increase in computational cost and latency. This makes them challenging to apply to applications that demand very low latency, such as chatbots and virtual assistants. We propose Possibility Exploration Fine-Tuning (PEFT), a task-agnostic framework that enhances the text diversity of LLMs without increasing latency or computational cost. Given the same prompt, models fine-tuned with PEFT can simultaneously generate multiple diverse responses, each corresponding with a controllable possibility number. Experiments on dialogue and story generation tasks demonstrate that PEFT significantly enhances the diversity of LLM outputs, as evidenced by lower similarity between candidate responses. Since PEFT emphasizes semantic diversity over lexical diversity, it can also notably reduce demographic bias in dialogue systems. The implementations and datasets are available in our repository: https://github.com/mailong25/peft_diversity

Juan Leon Alcazar, Long Mai, Federico Perazzi et al.

Humans are arguably one of the most important subjects in video streams, many real-world applications such as video summarization or video editing workflows often require the automatic search and retrieval of a person of interest. Despite tremendous efforts in the person reidentification and retrieval domains, few works have developed audiovisual search strategies. In this paper, we present the Audiovisual Person Search dataset (APES), a new dataset composed of untrimmed videos whose audio (voices) and visual (faces) streams are densely annotated. APES contains over 1.9K identities labeled along 36 hours of video, making it the largest dataset available for untrimmed audiovisual person search. A key property of APES is that it includes dense temporal annotations that link faces to speech segments of the same identity. To showcase the potential of our new dataset, we propose an audiovisual baseline and benchmark for person retrieval. Our study shows that modeling audiovisual cues benefits the recognition of people's identities. To enable reproducibility and promote future research, the dataset annotations and baseline code are available at: https://github.com/fuankarion/audiovisual-person-search

Wei Yin, Jianming Zhang, Oliver Wang et al.

Despite significant progress in monocular depth estimation in the wild, recent state-of-the-art methods cannot be used to recover accurate 3D scene shape due to an unknown depth shift induced by shift-invariant reconstruction losses used in mixed-data depth prediction training, and possible unknown camera focal length. We investigate this problem in detail, and propose a two-stage framework that first predicts depth up to an unknown scale and shift from a single monocular image, and then use 3D point cloud encoders to predict the missing depth shift and focal length that allow us to recover a realistic 3D scene shape. In addition, we propose an image-level normalized regression loss and a normal-based geometry loss to enhance depth prediction models trained on mixed datasets. We test our depth model on nine unseen datasets and achieve state-of-the-art performance on zero-shot dataset generalization. Code is available at: https://git.io/Depth

Jinbo Xing, Long Mai, Cusuh Ham et al.

This paper presents a method that allows users to design cinematic video shots in the context of image-to-video generation. Shot design, a critical aspect of filmmaking, involves meticulously planning both camera movements and object motions in a scene. However, enabling intuitive shot design in modern image-to-video generation systems presents two main challenges: first, effectively capturing user intentions on the motion design, where both camera movements and scene-space object motions must be specified jointly; and second, representing motion information that can be effectively utilized by a video diffusion model to synthesize the image animations. To address these challenges, we introduce MotionCanvas, a method that integrates user-driven controls into image-to-video (I2V) generation models, allowing users to control both object and camera motions in a scene-aware manner. By connecting insights from classical computer graphics and contemporary video generation techniques, we demonstrate the ability to achieve 3D-aware motion control in I2V synthesis without requiring costly 3D-related training data. MotionCanvas enables users to intuitively depict scene-space motion intentions, and translates them into spatiotemporal motion-conditioning signals for video diffusion models. We demonstrate the effectiveness of our method on a wide range of real-world image content and shot-design scenarios, highlighting its potential to enhance the creative workflows in digital content creation and adapt to various image and video editing applications.

Andrew Bond, Jui-Hsien Wang, Long Mai et al.

Efficient neural representations for dynamic video scenes are critical for applications ranging from video compression to interactive simulations. Yet, existing methods often face challenges related to high memory usage, lengthy training times, and temporal consistency. To address these issues, we introduce a novel neural video representation that combines 3D Gaussian splatting with continuous camera motion modeling. By leveraging Neural ODEs, our approach learns smooth camera trajectories while maintaining an explicit 3D scene representation through Gaussians. Additionally, we introduce a spatiotemporal hierarchical learning strategy, progressively refining spatial and temporal features to enhance reconstruction quality and accelerate convergence. This memory-efficient approach achieves high-quality rendering at impressive speeds. Experimental results show that our hierarchical learning, combined with robust camera motion modeling, captures complex dynamic scenes with strong temporal consistency, achieving state-of-the-art performance across diverse video datasets in both high- and low-motion scenarios.

Quynh Phung, Long Mai, Fabian David Caba Heilbron et al.

We present CineVerse, a novel framework for the task of cinematic scene composition. Similar to traditional multi-shot generation, our task emphasizes the need for consistency and continuity across frames. However, our task also focuses on addressing challenges inherent to filmmaking, such as multiple characters, complex interactions, and visual cinematic effects. In order to learn to generate such content, we first create the CineVerse dataset. We use this dataset to train our proposed two-stage approach. First, we prompt a large language model (LLM) with task-specific instructions to take in a high-level scene description and generate a detailed plan for the overall setting and characters, as well as the individual shots. Then, we fine-tune a text-to-image generation model to synthesize high-quality visual keyframes. Experimental results demonstrate that CineVerse yields promising improvements in generating visually coherent and contextually rich movie scenes, paving the way for further exploration in cinematic video synthesis.

Hung Huy Nguyen, Pooyan Rahmanzadehgervi, Long Mai et al.

Detecting object-level changes between two images across possibly different views is a core task in many applications that involve visual inspection or camera surveillance. Existing change-detection approaches suffer from three major limitations: (1) lack of evaluation on image pairs that contain no changes, leading to unreported false positive rates; (2) lack of correspondences (i.e., localizing the regions before and after a change); and (3) poor zero-shot generalization across different domains. To address these issues, we introduce a novel method that leverages change correspondences (a) during training to improve change detection accuracy, and (b) at test time, to minimize false positives. That is, we harness the supervision labels of where an object is added or removed to supervise change detectors, improving their accuracy over previous work by a large margin. Our work is also the first to predict correspondences between pairs of detected changes using estimated homography and the Hungarian algorithm. Our model demonstrates superior performance over existing methods, achieving state-of-the-art results in change detection and change correspondence accuracy across both in-distribution and zero-shot benchmarks.

Yuan Yao, Yicong Hong, Difan Liu et al.

The quadratic computational complexity of self-attention in diffusion transformers (DiT) introduces substantial computational costs in high-resolution image generation. While the linear-complexity Mamba model emerges as a potential alternative, direct Mamba training remains empirically challenging. To address this issue, this paper introduces diffusion transformer-to-mamba distillation (T2MD), forming an efficient training pipeline that facilitates the transition from the self-attention-based transformer to the linear complexity state-space model Mamba. We establish a diffusion self-attention and Mamba hybrid model that simultaneously achieves efficiency and global dependencies. With the proposed layer-level teacher forcing and feature-based knowledge distillation, T2MD alleviates the training difficulty and high cost of a state space model from scratch. Starting from the distilled 512$\times$512 resolution base model, we push the generation towards 2048$\times$2048 images via lightweight adaptation and high-resolution fine-tuning. Experiments demonstrate that our training path leads to low overhead but high-quality text-to-image generation. Importantly, our results also justify the feasibility of using sequential and causal Mamba models for generating non-causal visual output, suggesting the potential for future exploration.

Yitian Zhang, Long Mai, Aniruddha Mahapatra et al.

We present a novel perspective on learning video embedders for generative modeling: rather than requiring an exact reproduction of an input video, an effective embedder should focus on synthesizing visually plausible reconstructions. This relaxed criterion enables substantial improvements in compression ratios without compromising the quality of downstream generative models. Specifically, we propose replacing the conventional encoder-decoder video embedder with an encoder-generator framework that employs a diffusion transformer (DiT) to synthesize missing details from a compact latent space. Therein, we develop a dedicated latent conditioning module to condition the DiT decoder on the encoded video latent embedding. Our experiments demonstrate that our approach enables superior encoding-decoding performance compared to state-of-the-art methods, particularly as the compression ratio increases. To demonstrate the efficacy of our approach, we report results from our video embedders achieving a temporal compression ratio of up to 32x (8x higher than leading video embedders) and validate the robustness of this ultra-compact latent space for text-to-video generation, providing a significant efficiency boost in latent diffusion model training and inference.

Yicong Hong, Long Mai, Yuan Yao et al.

While Transformers have become the dominant architecture for visual generation, linear attention models, such as the state-space models (SSM), are increasingly recognized for their efficiency in processing long visual sequences. However, the essential efficiency of these models comes from formulating a limited recurrent state, enforcing causality among tokens that are prone to inconsistent modeling of N-dimensional visual data, leaving questions on their capacity to generate long non-causal sequences. In this paper, we explore the boundary of SSM on image and video generation by building the largest-scale diffusion SSM-Transformer hybrid model to date (5B parameters) based on the sub-quadratic bi-directional Hydra and self-attention, and generate up to 2K images and 360p 8 seconds (16 FPS) videos. Our results demonstrate that the model can produce faithful results aligned with complex text prompts and temporal consistent videos with high dynamics, suggesting the great potential of using SSMs for visual generation tasks.

Aniruddha Mahapatra, Long Mai, David Bourgin et al.

Video tokenizers are essential for latent video diffusion models, converting raw video data into spatiotemporally compressed latent spaces for efficient training. However, extending state-of-the-art video tokenizers to achieve a temporal compression ratio beyond 4x without increasing channel capacity poses significant challenges. In this work, we propose an alternative approach to enhance temporal compression. We find that the reconstruction quality of temporally subsampled videos from a low-compression encoder surpasses that of high-compression encoders applied to original videos. This indicates that high-compression models can leverage representations from lower-compression models. Building on this insight, we develop a bootstrapped high-temporal-compression model that progressively trains high-compression blocks atop well-trained lower-compression models. Our method includes a cross-level feature-mixing module to retain information from the pretrained low-compression model and guide higher-compression blocks to capture the remaining details from the full video sequence. Evaluation of video benchmarks shows that our method significantly improves reconstruction quality while increasing temporal compression compared to directly training the full model. Furthermore, the resulting compact latent space effectively trains a video diffusion model for high-quality video generation with a significantly reduced token budget.

Pooyan Rahmanzadehgervi, Hung Huy Nguyen, Rosanne Liu et al.

Multi-head self-attention (MHSA) is a key component of Transformers, a widely popular architecture in both language and vision. Multiple heads intuitively enable different parallel processes over the same input. Yet, they also obscure the attribution of each input patch to the output of a model. We propose a novel 1-head Transformer Attention Bottleneck (TAB) layer, inserted after the traditional MHSA architecture, to serve as an attention bottleneck for interpretability and intervention. Unlike standard self-attention, TAB constrains the total attention over all patches to $\in [0, 1]$. That is, when the total attention is 0, no visual information is propagated further into the network, and the vision-language model (VLM) would default to a generic, image-independent response. To demonstrate the advantages of TAB, we train VLMs with TAB to perform image-difference captioning. Over three datasets, our models perform similarly to baseline VLMs in captioning but the bottleneck is superior in localizing changes and in identifying when no changes occur. TAB is the first architecture to enable users to debug by editing attention, which often produces expected outputs by VLMs.

Hanshu Yan, Jun Hao Liew, Long Mai et al.

This paper addresses the issue of modifying the visual appearance of videos while preserving their motion. A novel framework, named MagicProp, is proposed, which disentangles the video editing process into two stages: appearance editing and motion-aware appearance propagation. In the first stage, MagicProp selects a single frame from the input video and applies image-editing techniques to modify the content and/or style of the frame. The flexibility of these techniques enables the editing of arbitrary regions within the frame. In the second stage, MagicProp employs the edited frame as an appearance reference and generates the remaining frames using an autoregressive rendering approach. To achieve this, a diffusion-based conditional generation model, called PropDPM, is developed, which synthesizes the target frame by conditioning on the reference appearance, the target motion, and its previous appearance. The autoregressive editing approach ensures temporal consistency in the resulting videos. Overall, MagicProp combines the flexibility of image-editing techniques with the superior temporal consistency of autoregressive modeling, enabling flexible editing of object types and aesthetic styles in arbitrary regions of input videos while maintaining good temporal consistency across frames. Extensive experiments in various video editing scenarios demonstrate the effectiveness of MagicProp.

Thang M. Pham, Trung Bui, Long Mai et al.

A principle behind dozens of attribution methods is to take the prediction difference between before-and-after an input feature (here, a token) is removed as its attribution. A popular Input Marginalization (IM) method (Kim et al., 2020) uses BERT to replace a token, yielding more plausible counterfactuals. While Kim et al. (2020) reported that IM is effective, we find this conclusion not convincing as the DeletionBERT metric used in their paper is biased towards IM. Importantly, this bias exists in Deletion-based metrics, including Insertion, Sufficiency, and Comprehensiveness. Furthermore, our rigorous evaluation using 6 metrics and 3 datasets finds no evidence that IM is better than a Leave-One-Out (LOO) baseline. We find two reasons why IM is not better than LOO: (1) deleting a single word from the input only marginally reduces a classifier's accuracy; and (2) a highly predictable word is always given near-zero attribution, regardless of its true importance to the classifier. In contrast, making LIME samples more natural via BERT consistently improves LIME accuracy under several ROAR metrics.

Alexander Black, Tu Bui, Long Mai et al.

We present an algorithm for searching image collections using free-hand sketches that describe the appearance and relative positions of multiple objects. Sketch based image retrieval (SBIR) methods predominantly match queries containing a single, dominant object invariant to its position within an image. Our work exploits drawings as a concise and intuitive representation for specifying entire scene compositions. We train a convolutional neural network (CNN) to encode masked visual features from sketched objects, pooling these into a spatial descriptor encoding the spatial relationships and appearances of objects in the composition. Training the CNN backbone as a Siamese network under triplet loss yields a metric search embedding for measuring compositional similarity which may be efficiently leveraged for visual search by applying product quantization.

S. Mahdi H. Miangoleh, Sebastian Dille, Long Mai et al.

Neural networks have shown great abilities in estimating depth from a single image. However, the inferred depth maps are well below one-megapixel resolution and often lack fine-grained details, which limits their practicality. Our method builds on our analysis on how the input resolution and the scene structure affects depth estimation performance. We demonstrate that there is a trade-off between a consistent scene structure and the high-frequency details, and merge low- and high-resolution estimations to take advantage of this duality using a simple depth merging network. We present a double estimation method that improves the whole-image depth estimation and a patch selection method that adds local details to the final result. We demonstrate that by merging estimations at different resolutions with changing context, we can generate multi-megapixel depth maps with a high level of detail using a pre-trained model.

Thang M. Pham, Trung Bui, Long Mai et al.

Do state-of-the-art natural language understanding models care about word order - one of the most important characteristics of a sequence? Not always! We found 75% to 90% of the correct predictions of BERT-based classifiers, trained on many GLUE tasks, remain constant after input words are randomly shuffled. Despite BERT embeddings are famously contextual, the contribution of each individual word to downstream tasks is almost unchanged even after the word's context is shuffled. BERT-based models are able to exploit superficial cues (e.g. the sentiment of keywords in sentiment analysis; or the word-wise similarity between sequence-pair inputs in natural language inference) to make correct decisions when tokens are arranged in random orders. Encouraging classifiers to capture word order information improves the performance on most GLUE tasks, SQuAD 2.0 and out-of-samples. Our work suggests that many GLUE tasks are not challenging machines to understand the meaning of a sentence.

Simon Niklaus, Long Mai, Oliver Wang

Video frame interpolation, the synthesis of novel views in time, is an increasingly popular research direction with many new papers further advancing the state of the art. But as each new method comes with a host of variables that affect the interpolation quality, it can be hard to tell what is actually important for this task. In this work, we show, somewhat surprisingly, that it is possible to achieve near state-of-the-art results with an older, simpler approach, namely adaptive separable convolutions, by a subtle set of low level improvements. In doing so, we propose a number of intuitive but effective techniques to improve the frame interpolation quality, which also have the potential to other related applications of adaptive convolutions such as burst image denoising, joint image filtering, or video prediction.

Juan Leon Alcazar, Fabian Caba Heilbron, Long Mai et al.

Current methods for active speak er detection focus on modeling short-term audiovisual information from a single speaker. Although this strategy can be enough for addressing single-speaker scenarios, it prevents accurate detection when the task is to identify who of many candidate speakers are talking. This paper introduces the Active Speaker Context, a novel representation that models relationships between multiple speakers over long time horizons. Our Active Speaker Context is designed to learn pairwise and temporal relations from an structured ensemble of audio-visual observations. Our experiments show that a structured feature ensemble already benefits the active speaker detection performance. Moreover, we find that the proposed Active Speaker Context improves the state-of-the-art on the AVA-ActiveSpeaker dataset achieving a mAP of 87.1%. We present ablation studies that verify that this result is a direct consequence of our long-term multi-speaker analysis.

Zhuowan Li, Quan Tran, Long Mai et al.

While image captioning has progressed rapidly, existing works focus mainly on describing single images. In this paper, we introduce a new task, context-aware group captioning, which aims to describe a group of target images in the context of another group of related reference images. Context-aware group captioning requires not only summarizing information from both the target and reference image group but also contrasting between them. To solve this problem, we propose a framework combining self-attention mechanism with contrastive feature construction to effectively summarize common information from each image group while capturing discriminative information between them. To build the dataset for this task, we propose to group the images and generate the group captions based on single image captions using scene graphs matching. Our datasets are constructed on top of the public Conceptual Captions dataset and our new Stock Captions dataset. Experiments on the two datasets show the effectiveness of our method on this new task. Related Datasets and code are released at https://lizw14.github.io/project/groupcap .

Thu Nguyen-Phuoc, Christian Richardt, Long Mai et al.

We present BlockGAN, an image generative model that learns object-aware 3D scene representations directly from unlabelled 2D images. Current work on scene representation learning either ignores scene background or treats the whole scene as one object. Meanwhile, work that considers scene compositionality treats scene objects only as image patches or 2D layers with alpha maps. Inspired by the computer graphics pipeline, we design BlockGAN to learn to first generate 3D features of background and foreground objects, then combine them into 3D features for the wholes cene, and finally render them into realistic images. This allows BlockGAN to reason over occlusion and interaction between objects' appearance, such as shadow and lighting, and provides control over each object's 3D pose and identity, while maintaining image realism. BlockGAN is trained end-to-end, using only unlabelled single images, without the need for 3D geometry, pose labels, object masks, or multiple views of the same scene. Our experiments show that using explicit 3D features to represent objects allows BlockGAN to learn disentangled representations both in terms of objects (foreground and background) and their properties (pose and identity).

Qi Li, Long Mai, Michael A. Alcorn et al.

Large, pre-trained generative models have been increasingly popular and useful to both the research and wider communities. Specifically, BigGANs a class-conditional Generative Adversarial Networks trained on ImageNet---achieved excellent, state-of-the-art capability in generating realistic photos. However, fine-tuning or training BigGANs from scratch is practically impossible for most researchers and engineers because (1) GAN training is often unstable and suffering from mode-collapse; and (2) the training requires a significant amount of computation, 256 Google TPUs for 2 days or 8xV100 GPUs for 15 days. Importantly, many pre-trained generative models both in NLP and image domains were found to contain biases that are harmful to society. Thus, we need computationally-feasible methods for modifying and re-purposing these huge, pre-trained models for downstream tasks. In this paper, we propose a cost-effective optimization method for improving and re-purposing BigGANs by fine-tuning only the class-embedding layer. We show the effectiveness of our model-editing approach in three tasks: (1) significantly improving the realism and diversity of samples of complete mode-collapse classes; (2) re-purposing ImageNet BigGANs for generating images for Places365; and (3) de-biasing or improving the sample diversity for selected ImageNet classes.

Haotian Zhang, Long Mai, Ning Xu et al.

We propose a novel video inpainting algorithm that simultaneously hallucinates missing appearance and motion (optical flow) information, building upon the recent 'Deep Image Prior' (DIP) that exploits convolutional network architectures to enforce plausible texture in static images. In extending DIP to video we make two important contributions. First, we show that coherent video inpainting is possible without a priori training. We take a generative approach to inpainting based on internal (within-video) learning without reliance upon an external corpus of visual data to train a one-size-fits-all model for the large space of general videos. Second, we show that such a framework can jointly generate both appearance and flow, whilst exploiting these complementary modalities to ensure mutual consistency. We show that leveraging appearance statistics specific to each video achieves visually plausible results whilst handling the challenging problem of long-term consistency.

Simon Niklaus, Long Mai, Jimei Yang et al.

The Ken Burns effect allows animating still images with a virtual camera scan and zoom. Adding parallax, which results in the 3D Ken Burns effect, enables significantly more compelling results. Creating such effects manually is time-consuming and demands sophisticated editing skills. Existing automatic methods, however, require multiple input images from varying viewpoints. In this paper, we introduce a framework that synthesizes the 3D Ken Burns effect from a single image, supporting both a fully automatic mode and an interactive mode with the user controlling the camera. Our framework first leverages a depth prediction pipeline, which estimates scene depth that is suitable for view synthesis tasks. To address the limitations of existing depth estimation methods such as geometric distortions, semantic distortions, and inaccurate depth boundaries, we develop a semantic-aware neural network for depth prediction, couple its estimate with a segmentation-based depth adjustment process, and employ a refinement neural network that facilitates accurate depth predictions at object boundaries. According to this depth estimate, our framework then maps the input image to a point cloud and synthesizes the resulting video frames by rendering the point cloud from the corresponding camera positions. To address disocclusions while maintaining geometrically and temporally coherent synthesis results, we utilize context-aware color- and depth-inpainting to fill in the missing information in the extreme views of the camera path, thus extending the scene geometry of the point cloud. Experiments with a wide variety of image content show that our method enables realistic synthesis results. Our study demonstrates that our system allows users to achieve better results while requiring little effort compared to existing solutions for the 3D Ken Burns effect creation.

Peng Zhou, Long Mai, Jianming Zhang et al.

Incremental learning targets at achieving good performance on new categories without forgetting old ones. Knowledge distillation has been shown critical in preserving the performance on old classes. Conventional methods, however, sequentially distill knowledge only from the last model, leading to performance degradation on the old classes in later incremental learning steps. In this paper, we propose a multi-model and multi-level knowledge distillation strategy. Instead of sequentially distilling knowledge only from the last model, we directly leverage all previous model snapshots. In addition, we incorporate an auxiliary distillation to further preserve knowledge encoded at the intermediate feature levels. To make the model more memory efficient, we adapt mask based pruning to reconstruct all previous models with a small memory footprint. Experiments on standard incremental learning benchmarks show that our method preserves the knowledge on old classes better and improves the overall performance over standard distillation techniques.

Michael A. Alcorn, Qi Li, Zhitao Gong et al.

Despite excellent performance on stationary test sets, deep neural networks (DNNs) can fail to generalize to out-of-distribution (OoD) inputs, including natural, non-adversarial ones, which are common in real-world settings. In this paper, we present a framework for discovering DNN failures that harnesses 3D renderers and 3D models. That is, we estimate the parameters of a 3D renderer that cause a target DNN to misbehave in response to the rendered image. Using our framework and a self-assembled dataset of 3D objects, we investigate the vulnerability of DNNs to OoD poses of well-known objects in ImageNet. For objects that are readily recognized by DNNs in their canonical poses, DNNs incorrectly classify 97% of their pose space. In addition, DNNs are highly sensitive to slight pose perturbations. Importantly, adversarial poses transfer across models and datasets. We find that 99.9% and 99.4% of the poses misclassified by Inception-v3 also transfer to the AlexNet and ResNet-50 image classifiers trained on the same ImageNet dataset, respectively, and 75.5% transfer to the YOLOv3 object detector trained on MS COCO.

Simon Niklaus, Long Mai, Feng Liu

Standard video frame interpolation methods first estimate optical flow between input frames and then synthesize an intermediate frame guided by motion. Recent approaches merge these two steps into a single convolution process by convolving input frames with spatially adaptive kernels that account for motion and re-sampling simultaneously. These methods require large kernels to handle large motion, which limits the number of pixels whose kernels can be estimated at once due to the large memory demand. To address this problem, this paper formulates frame interpolation as local separable convolution over input frames using pairs of 1D kernels. Compared to regular 2D kernels, the 1D kernels require significantly fewer parameters to be estimated. Our method develops a deep fully convolutional neural network that takes two input frames and estimates pairs of 1D kernels for all pixels simultaneously. Since our method is able to estimate kernels and synthesizes the whole video frame at once, it allows for the incorporation of perceptual loss to train the neural network to produce visually pleasing frames. This deep neural network is trained end-to-end using widely available video data without any human annotation. Both qualitative and quantitative experiments show that our method provides a practical solution to high-quality video frame interpolation.

Simon Niklaus, Long Mai, Feng Liu

Video frame interpolation typically involves two steps: motion estimation and pixel synthesis. Such a two-step approach heavily depends on the quality of motion estimation. This paper presents a robust video frame interpolation method that combines these two steps into a single process. Specifically, our method considers pixel synthesis for the interpolated frame as local convolution over two input frames. The convolution kernel captures both the local motion between the input frames and the coefficients for pixel synthesis. Our method employs a deep fully convolutional neural network to estimate a spatially-adaptive convolution kernel for each pixel. This deep neural network can be directly trained end to end using widely available video data without any difficult-to-obtain ground-truth data like optical flow. Our experiments show that the formulation of video interpolation as a single convolution process allows our method to gracefully handle challenges like occlusion, blur, and abrupt brightness change and enables high-quality video frame interpolation.