ICCV 2021 Papers

In this paper, we look at the problem of few-shot image classification that aims to learn a classifier for previously unseen classes and domains from few labeled samples. Recent methods use various adaptation strategies for aligning their visual representations to new domains or select the relevant ones from multiple domain-specific feature extractors. In this work, we present URL, which learns a single set of universal visual representations by distilling knowledge of multiple domain-specific networks after co-aligning their features with the help of adapters and centered kernel alignment. We show that the universal representations can be further refined for previously unseen domains by an efficient adaptation step in a similar spirit to distance learning methods. We rigorously evaluate our model in the recent Meta-Dataset benchmark and demonstrate that it significantly outperforms the previous methods while being more efficient.

In this work, for the first time, we address the problem of universal cross-domain retrieval, where the test data can belong to classes or domains which are unseen during training. Due to dynamically increasing number of categories and practical constraint of training on every possible domain, which requires large amounts of data, generalizing to both unseen classes and domains is important. Towards that goal, we propose SnMpNet (Semantic Neighbourhood and Mixture Prediction Network), which incorporates two novel losses to account for the unseen classes and domains encountered during testing. Specifically, we introduce a novel Semantic Neighborhood loss to bridge the knowledge gap between seen and unseen classes and ensure that the latent space embedding of the unseen classes is semantically meaningful with respect to its neighboring classes. We also introduce a mix-up based supervision at image-level as well as semantic-level of the data for training with the Mixture Prediction loss, which helps in efficient retrieval when the query belongs to an unseen domain. These losses are incorporated on the SE-ResNet50 backbone to obtain SnMpNet. Extensive experiments on two large-scale datasets, Sketchy Extended and DomainNet, and thorough comparisons with state-of-the-art justify the effectiveness of the proposed model.

The nonlocal-based blocks are designed for capturing long-range spatial-temporal dependencies in computer vision tasks. Although having shown excellent performance, they still lack the mechanism to encode the rich, structured information among elements in an image or video. In this paper, to theoretically analyze the property of these nonlocal-based blocks, we provide a new perspective to interpret them, where we view them as a set of graph filters generated on a fully-connected graph. Specifically, when choosing the Chebyshev graph filter, a unified formulation can be derived for explaining and analyzing the existing nonlocal-based blocks (e.g., nonlocal block, nonlocal stage, double attention block). Furthermore, by concerning the property of spectral, we propose an efficient and robust spectral nonlocal block, which can be more robust and flexible to catch long-range dependencies when inserted into deep neural networks than the existing nonlocal blocks. Experimental results demonstrate the clear-cut improvements and practical applicabilities of our method on image classification, action recognition, semantic segmentation, and person re-identification tasks. Code are available at https://github.com/zh460045050/SNL_ICCV2021.

Recently, the problem of inaccurate learning targets in crowd counting draws increasing attention. Inspired by a few pioneering work, we solve this problem by trying to predict the indices of pre-defined interval bins of counts instead of the count values themselves. However, an inappropriate interval setting might make the count error contributions from different intervals extremely imbalanced, leading to inferior counting performance. Therefore, we propose a novel count interval partition criterion called Uniform Error Partition (UEP), which always keeps the expected counting error contributions equal for all intervals to minimize the prediction risk. Then to mitigate the inevitably introduced discretization errors in the count quantization process, we propose another criterion called Mean Count Proxies (MCP). The MCP criterion selects the best count proxy for each interval to represent its count value during inference, making the overall expected discretization error of an image nearly negligible. As far as we are aware, this work is the first to delve into such a classification task and ends up with a promising solution for count interval partition. Following the above two theoretically demonstrated criterions, we propose a simple yet effective model termed Uniform Error Partition Network (UEPNet), which achieves state-of-the-art performance on several challenging datasets. The codes will be available at: https://github.com/TencentYoutuResearch/CrowdCounting-UEPNet.

Building an interactive artificial intelligence that can ask questions about the real world is one of the biggest challenges for vision and language problems. In particular, goal-oriented visual dialogue, where the aim of the agent is to seek information by asking questions during a turn-taking dialogue, has been gaining scholarly attention recently. While several existing models based on the GuessWhat?! dataset have been proposed, the Questioner typically asks simple category-based questions or absolute spatial questions. This might be problematic for complex scenes where the objects share attributes or in cases where descriptive questions are required to distinguish objects. In this paper, we propose a novel Questioner architecture, called Unified Questioner Transformer (UniQer), for descriptive question generation with referring expressions. In addition, we build a goal-oriented visual dialogue task called CLEVR Ask. It synthesizes complex scenes that require the Questioner to generate descriptive questions. We train our model with two variants of CLEVR Ask datasets. The results of the quantitative and qualitative evaluations show that UniQer outperforms the baseline.

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

Current state-of-the-art object detection and segmentation methods work well under the closed-world assumption. This closed-world setting assumes that the list of object categories is available during training and deployment. However, many real-world applications require detecting or segmenting novel objects, i.e., object categories never seen during training. In this paper, we present, UVO (Unidentified Video Objects), a new benchmark for open-world class-agnostic object segmentation in videos. Besides shifting the focus to the open-world setup, UVO is significantly larger, providing approximately 6 times more videos compared with DAVIS, and 7 times more mask (instance) annotations per video compared with YouTube-VO(I)S. UVO is also more challenging as it includes many videos with crowded scenes and complex background motions. We also demonstrated that UVO can be used for other applications, such as object tracking and super-voxel segmentation. We believe that UVO is a versatile testbed for researchers to develop novel approaches for open-world class-agnostic object segmentation, and inspires new research directions towards a more comprehensive video understanding beyond classification and detection.

We propose UniT, a Unified Transformer model to simultaneously learn the most prominent tasks across different domains, ranging from object detection to natural language understanding and multimodal reasoning. Based on the transformer encoder-decoder architecture, our UniT model encodes each input modality with an encoder and makes predictions on each task with a shared decoder over the encoded input representations, followed by task-specific output heads. The entire model is jointly trained end-to-end with losses from each task. Compared to previous efforts on multi-task learning with transformers, we share the same model parameters across all tasks instead of separately fine-tuning task-specific models and handle a much higher variety of tasks across different domains. In our experiments, we learn 7 tasks jointly over 8 datasets, achieving strong performance on each task with significantly fewer parameters. Our code is available in MMF at https://mmf.sh.

The performance of a computer vision model depends on the size and quality of its training data. Recent studies have unveiled previously-unknown composition biases in common image datasets which then lead to skewed model outputs, and have proposed methods to mitigate these biases. However, most existing works assume that human-generated annotations can be considered gold-standard and unbiased. In this paper, we reveal that this assumption can be problematic, and that special care should be taken to prevent models from learning such annotation biases. We focus on facial expression recognition and compare the label biases between lab-controlled and in-the-wild datasets. We demonstrate that many expression datasets contain significant annotation biases between genders, especially when it comes to the happy and angry expressions, and that traditional methods cannot fully mitigate such biases in trained models. To remove expression annotation bias, we propose an AU-Calibrated Facial Expression Recognition (AUC-FER) framework that utilizes facial action units (AUs) and incorporates the triplet loss into the objective function. Experimental results suggest that the proposed method is more effective in removing expression annotation bias than existing techniques.

Image captioning is an important task for benchmarking visual reasoning and for enabling accessibility for people with vision impairments. However, as in many machine learning settings, social biases can influence image captioning in undesirable ways. In this work, we study bias propagation pathways within image captioning, focusing specifically on the COCO dataset. Prior work has analyzed gender bias in captions using automatically-derived gender labels; here we examine racial and intersectional biases using manual annotations. Our first contribution is in annotating the perceived gender and skin color of 28,315 of the depicted people after obtaining IRB approval. Using these annotations, we compare racial biases present in both manual and automatically-generated image captions. We demonstrate differences in caption performance, sentiment, and word choice between images of lighter versus darker-skinned people. Further, we find the magnitude of these differences to be greater in modern captioning systems compared to older ones, thus leading to concerns that without proper consideration and mitigation these differences will only become increasingly prevalent. Code and data is available at https://princetonvisualai.github.io/imagecaptioning-bias/.

Deep Convolutional Neural Networks (CNNs) have long been the architecture of choice for computer vision tasks. Recently, Transformer-based architectures like Vision Transformer (ViT) have matched or even surpassed ResNets for image classification. However, details of the Transformer architecture such as the use of non-overlapping patches lead one to wonder whether these networks are as robust. In this paper, we perform an extensive study of a variety of different measures of robustness of ViT models and compare the findings to ResNet baselines. We investigate robustness to input perturbations as well as robustness to model perturbations. We find that when pre-trained with a sufficient amount of data, ViT models are at least as robust as the ResNet counterparts on a broad range of perturbations. We also find that Transformers are robust to the removal of almost any single layer, and that while activations from later layers are highly correlated with each other, they nevertheless play an important role in classification.

We tackle the challenge of learning a distribution over complex, realistic, indoor scenes. In this paper, we introduce Generative Scene Networks (GSN), which learns to decompose scenes into a collection of many local radiance fields that can be rendered from a free moving camera. Our model can be used as a prior to generate new scenes, or to complete a scene given only sparse 2D observations. Recent work has shown that generative models of radiance fields can capture properties such as multi-view consistency and view-dependent lighting. However, these models are specialized for constrained viewing of single objects, such as cars or faces. Due to the size and complexity of realistic indoor environments, existing models lack the representational capacity to adequately capture them. Our decomposition scheme scales to larger and more complex scenes while preserving details and diversity, and the learned prior enables high-quality rendering from view-points that are significantly different from observed viewpoints. When compared to existing models, GSN produces quantitatively higher quality scene renderings across several different scene datasets.

Despite recent advancements in single-domain or single-object image generation, it is still challenging to generate complex scenes containing diverse, multiple objects and their interactions. Scene graphs, composed of nodes as objects and directed-edges as relationships among objects, offer an alternative representation of a scene that is more semantically grounded than images. We hypothesize that a generative model for scene graphs might be able to learn the underlying semantic structure of real-world scenes more effectively than images, and hence, generate realistic novel scenes in the form of scene graphs. In this work, we explore a new task for the unconditional generation of semantic scene graphs. We develop a deep auto-regressive model called SceneGraphGen which can directly learn the probability distribution over labelled and directed graphs using a hierarchical recurrent architecture. The model takes a seed object as input and generates a scene graph in a sequence of steps, each step generating an object node, followed by a sequence of relationship edges connecting to the previous nodes. We show that the scene graphs generated by SceneGraphGen are diverse and follow the semantic patterns of real-world scenes. Additionally, we demonstrate the application of the generated graphs in image synthesis, anomaly detection and scene graph completion.

Spotting objects that are visually adapted to their surroundings is challenging for both humans and AI. Conventional generic / salient object detection techniques are suboptimal for this task because they tend to only discover easy and clear objects, while overlooking the difficult-to-detect ones with inherent uncertainties derived from indistinguishable textures. In this work, we contribute a novel approach using a probabilistic representational model in combination with transformers to explicitly reason under uncertainties, namely uncertainty-guided transformer reasoning (UGTR), for camouflaged object detection. The core idea is to first learn a conditional distribution over the backbone's output to obtain initial estimates and associated uncertainties, and then reason over these uncertain regions with attention mechanism to produce final predictions. Our approach combines the benefits of both Bayesian learning and Transformer-based reasoning, allowing the model to handle camouflaged object detection by leveraging both deterministic and probabilistic information. We empirically demonstrate that our proposed approach can achieve higher accuracy than existing state-of-the-art models on CHAMELEON, CAMO and COD10K datasets. Code is available at https://github.com/fanyang587/UGTR.

Unsupervised domain adaptation for semantic segmentation aims to assign the pixel-level labels for unlabeled target domain by transferring knowledge from the labeled source domain. A typical self-supervised learning approach generates pseudo labels from the source model and then re-trains the model to fit the target distribution. However, it suffers from noisy pseudo labels due to the existence of domain shift. Related works alleviate this problem by selecting high-confidence predictions, but uncertain classes with low confidence scores have rarely been considered. This informative uncertainty is essential to enhance feature representation and align source and target domains. In this paper, we propose a novel uncertainty-aware pseudo label refinery framework considering two crucial factors simultaneously. First, we progressively enhance the feature alignment model via the target-guided uncertainty rectifying framework. Second, we provide an uncertainty-aware pseudo label assignment strategy without any manually designed threshold to reduce the noisy labels. Extensive experiments demonstrate the effectiveness of our proposed approach and achieve state-of-the-art performance on two standard synthetic-2-real tasks.

Despite the recent success of 3D human reconstruction methods, recovering the accurate and smooth 3D human motion from video is still challenging. Designing a temporal model in the encoding stage is not sufficient enough to settle the trade-off problem between the per-frame accuracy and the motion smoothness. To address this problem, we approach some of the fundamental problems of 3D reconstruction tasks, simultaneously predicting 3D pose and 3D motion dynamics. First, we utilize the power of uncertainty to address the problem of multiple 3D configurations resulting in the same 2D projections. Second, we confirmed that dividing the body into local regions shows outstanding results for estimating 3D motion dynamics. In this paper, we propose (i) an encoder that makes two different estimations: a static feature that presents 2D pose feature as distribution and a dynamic feature that includes optical flow information and (ii) a decoder that divides the body into five different local regions to estimate the 3D motion dynamics of each region. We demonstrate how our method recovers the accurate and smooth motion and achieves the state-of-the-art results for both constrained and in-the-wild videos.

Unsupervised image-to-image translation aims at learning the mapping from the source to target domain without using paired images for training. An essential yet restrictive assumption for unsupervised image translation is that the two domains are aligned, e.g., for the selfie2anime task, the anime (selfie) domain must contain only anime (selfie) face images that can be translated to some images in the other domain. Collecting aligned domains can be laborious and needs lots of attention. In this paper, we consider the task of image translation between two unaligned domains, which may arise for various possible reasons. To solve this problem, we propose to select images based on importance reweighting and develop a method to learn the weights and perform translation simultaneously and automatically. We compare the proposed method with state-of-the-art image translation approaches and present qualitative and quantitative results on different tasks with unaligned domains. Extensive empirical evidence demonstrates the usefulness of the proposed problem formulation and the superiority of our method.

Recovering dense human poses from images plays a critical role in establishing an image-to-surface correspondence between RGB images and the 3D surface of the human body, serving the foundation of rich real-world applications, such as virtual humans, monocular-to-3d reconstruction. However, the popular DensePose-COCO dataset relies on a sophisticated manual annotation system, leading to severe limitations in acquiring the denser and more accurate annotated pose resources. In this work, we introduce a new 3D human-body model with a series of decoupled parameters that could freely control the generation of the body. Furthermore, we build a data generation system based on this decoupling 3D model, and construct an ultra dense synthetic benchmark UltraPose, containing around 1.3 billion corresponding points. Compared to the existing manually annotated DensePose-COCO dataset, the synthetic UltraPose has ultra dense image-to-surface correspondences without annotation cost and error. Our proposed UltraPose provides the largest benchmark and data resources for lifting the model capability in predicting more accurate dense poses. To promote future researches in this field, we also propose a transformer-based method to model the dense correspondence between 2D and 3D worlds. The proposed model trained on synthetic UltraPose can be applied to real-world scenarios, indicating the effectiveness of our benchmark and model.

Existing single image high dynamic range (HDR) reconstruction attempt to expand the range of luminance. They are not effective to generate plausible textures and colors in the reconstructed results, especially for high-density pixels in ultra-high-definition (UHD) images.To address these problems, we propose a new HDR reconstruction network for UHD images by collaboratively learning color and texture details. First, we propose a dual-path network to extract content and chromatic features at a reduced resolution of the low dynamic range (LDR) input. These two types features are used to fit bilatera-space affine models for real-time HDR reconstruction. To extract the main data structure of the LDR input, we propose to use 3D Tucker decomposition and reconstruction to prevents false edges and noise amplification in the learned bilateral grid. As a result, the high-quality content and chromatic features can be reconstructed capitalized on guided bilateral upsampling. Finally, we fuse these two full-resolution feature maps into the HDR reconstructed results.Our proposed method can achieve real-time processing for UHD image (about 160 fps).Experimental results demonstrate that the proposed algorithm performs favorably against the state-of-the-art HDR reconstruction approaches on public benchmarks and real-world UHD images.