ICCV 2021 Papers

This work addresses the problem of discovering, in an unsupervised manner, interpretable paths in the latent space of pretrained GANs, so as to provide an intuitive and easy way of controlling the underlying generative factors. In doing so, it addresses some of the limitations of the state-of-the-art works, namely, a) that they discover directions that are independent of the latent code, i.e., paths that are linear, and b) that their evaluation relies either on visual inspection or on laborious human labeling. More specifically, we propose to learn non-linear warpings on the latent space, each one parametrized by a set of RBF-based latent space warping functions, and where each warping gives rise to a family of non-linear paths via the gradient of the function. Building on the work of Voynov and Babenko, that discovers linear paths, we optimize the trainable parameters of the set of RBFs, so as that images that are generated by codes along different paths, are easily distinguishable by a discriminator network. This leads to easily distinguishable image transformations, such as pose and facial expressions in facial images. We show that linear paths can be derived as a special case of our method, and show experimentally that non-linear paths in the latent space lead to steeper, more disentangled and interpretable changes in the image space than in state-of-the art methods, both qualitatively and quantitatively. We make the code and the pretrained models publicly available at: https://github.com/chi0tzp/WarpedGANSpace.

Deep learning models for semantic segmentation rely on expensive, large-scale, manually annotated datasets. Labelling is a tedious process that can take hours per image. Automatically annotating video sequences by propagating sparsely labeled frames through time is a more scalable alternative. In this work, we propose a novel label propagation method, termed Warp-Refine Propagation, that combines semantic cues with geometric cues to efficiently auto-label videos. Our method learns to refine geometrically-warped labels and infuse them with learned semantic priors in a semi-supervised setting by leveraging cycle consistency across time. We quantitatively show that our method improves label-propagation by a noteworthy margin of 13.1 mIoU on the ApolloScape dataset. Furthermore, by training with the auto-labelled frames, we achieve competitive results on three semantic-segmentation benchmarks, improving the state-of-the-art by a large margin of 1.8 and 3.61 mIoU on NYU-V2 and KITTI, while matching the current best results on Cityscapes.

The key challenge in learning dense correspondences lies in the lack of ground-truth matches for real image pairs. While photometric consistency losses provide unsupervised alternatives, they struggle with large appearance changes, which are ubiquitous in geometric and semantic matching tasks. Moreover, methods relying on synthetic training pairs often suffer from poor generalisation to real data. We propose Warp Consistency, an unsupervised learning objective for dense correspondence regression. Our objective is effective even in settings with large appearance and view-point changes. Given a pair of real images, we first construct an image triplet by applying a randomly sampled warp to one of the original images. We derive and analyze all flow-consistency constraints arising between the triplet. From our observations and empirical results, we design a general unsupervised objective employing two of the derived constraints. We validate our warp consistency loss by training three recent dense correspondence networks for the geometric and semantic matching tasks. Our approach sets a new state-of-the-art on several challenging benchmarks, including MegaDepth, RobotCar and TSS. Code and models are at github.com/PruneTruong/DenseMatching.

Online continual learning from data streams in dynamic environments is a critical direction in the computer vision field. However, realistic benchmarks and fundamental studies in this line are still missing. To bridge the gap, we present a new online continual object detection benchmark with an egocentric video dataset, Objects Around Krishna (OAK). OAK adopts the KrishnaCAM videos, an ego-centric video stream collected over nine months by a graduate student. OAK provides exhaustive bounding box annotations of 80 video snippets (~17.5 hours) for 105 object categories in outdoor scenes. The emergence of new object categories in our benchmark follows a pattern similar to what a single person might see in their day-to-day life. The dataset also captures the natural distribution shifts as the person travels to different places. These egocentric long running videos provide a realistic playground for continual learning algorithms, especially in online embodied settings. We also introduce new evaluation metrics to evaluate the model performance and catastrophic forgetting and provide baseline studies for online continual object detection. We believe this benchmark will pose new exciting challenges for learning from non-stationary data in continual learning. The OAK dataset and the associated benchmark are released at https://oakdata.github.io/.

Discrete point cloud objects lack sufficient shape descriptors of 3D geometries. In this paper, we present a novel method for aggregating hypothetical curves in point clouds. Sequences of connected points (curves) are initially grouped by taking guided walks in the point clouds, and then subsequently aggregated back to augment their point-wise features. We provide an effective implementation of the proposed aggregation strategy including a novel curve grouping operator followed by a curve aggregation operator. Our method was benchmarked on several point cloud analysis tasks where we achieved the state-of-the-art classification accuracy of 94.2% on the ModelNet40 classification task, instance IoU of 86.8% on the ShapeNetPart segmentation task and cosine error of 0.11 on the ModelNet40 normal estimation task.

Transformer-based detector is a new paradigm in object detection, which aims to achieve pretty-well performance while eliminates the priori knowledge driven components, e.g., anchors, proposals and the NMS. DETR, the state-of-the-art model among them, is composed of three sub-modules, i.e., a CNN-based backbone and paired transformer encoder-decoder. The CNN is applied to extract local features and the transformer is used to capture global contexts. This pipeline, however, is not concise enough. In this paper, we propose WB-DETR (DETR-based detector Without Backbone) to prove that the reliance on CNN features extraction for a transformer-based detector is not necessary. Unlike the original DETR, WB-DETR is composed of only an encoder and a decoder without CNN backbone. For an input image, WB-DETR serializes it directly to encode the local features into each individual token. To make up the deficiency of transformer in modeling local information, we design an LIE-T2T (local information enhancement tokens to token) module to enhance the internal information of tokens after unfolding. Experimental results demonstrate that WB-DETR, the first pure-transformer detector without CNN to our knowledge, yields on par accuracy and faster inference speed with only half number of parameters compared with DETR baseline.

Deep learning technique has yielded significant improvements in point cloud completion with the aim of completing missing object shapes from partial inputs. However, most existing methods fail to recover realistic structures due to over-smoothing of fine-grained details. In this paper, we develop a voxel-based network for point cloud completion by leveraging edge generation (VE-PCN). We first embed point clouds into regular voxel grids, and then generate complete objects with the help of the hallucinated shape edges. This decoupled architecture together with a multi-scale grid feature learning is able to generate more realistic on-surface details. We evaluate our model on the publicly available completion datasets and show that it outperforms existing state-of-the-art approaches quantitatively and qualitatively. Our source code is available at https://github.com/xiaogangw/VE-PCN.

We present Voxel Transformer (VoTr), a novel and effective voxel-based Transformer backbone for 3D object detection from point clouds. Conventional 3D convolutional backbones in voxel-based 3D detectors cannot efficiently capture large context information, which is crucial for object recognition and localization, owing to the limited receptive fields. In this paper, we resolve the problem by introducing a Transformer-based architecture that enables long-range relationships between voxels by self-attention. Given the fact that non-empty voxels are naturally sparse but numerous, directly applying standard Transformer on voxels is non-trivial. To this end, we propose the sparse voxel module and the submanifold voxel module, which can operate on the empty and non-empty voxel positions effectively. To further enlarge the attention range while maintaining comparable computational overhead to the convolutional counterparts, we propose two attention mechanisms for multi-head attention in those two modules: Local Attention and Dilated Attention, and we further propose Fast Voxel Query to accelerate the querying process in multi-head attention. VoTr contains a series of sparse and submanifold voxel modules and can be applied in most voxel-based detectors. Our proposed VoTr shows consistent improvement over the convolutional baselines while maintaining computational efficiency on the KITTI dataset and the Waymo Open dataset.

To reconstruct a 3D scene from a set of calibrated views, traditional multi-view stereo techniques rely on two distinct stages: local depth maps computation and global depth maps fusion. Recent studies concentrate on deep neural architectures for depth estimation by using conventional depth fusion method or direct 3D reconstruction network by regressing Truncated Signed Distance Function (TSDF). In this paper, we advocate that replicating the traditional two stages framework with deep neural networks improves both the interpretability and the accuracy of the results. As mentioned, our network operates in two steps: 1) the local computation of the local depth maps with a deep MVS technique, and, 2) the depth maps and images' features fusion to build a single TSDF volume. In order to improve the matching performance between images acquired from very different viewpoints (e.g., large-baseline and rotations), we introduce a rotation-invariant 3D convolution kernel called PosedConv. The effectiveness of the proposed architecture is underlined via a large series of experiments conducted on the ScanNet dataset where our approach compares favorably against both traditional and deep learning techniques.

Diagnosing diseases from medical radiographs and writing reports requires professional knowledge and is time-consuming. To address this, automatic medical report generation approaches have recently gained interest. However, identifying diseases as well as correctly predicting their corresponding sizes, locations and other medical description patterns, which is essential for generating high-quality reports, is challenging. Although previous methods focused on producing readable reports, how to accurately detect and describe findings that match with the query X-Ray has not been successfully addressed. In this paper, we propose a multi-modality semantic attention model to integrate visual features, predicted key finding embeddings, as well as clinical features, and progressively decode reports with visual-textual semantic consistency. First, multi-modality features are extracted and attended with the hidden states from the sentence decoder, to encode enriched context vectors for better decoding a report. These modalities include regional visual features of scans, semantic word embeddings of the top-K findings predicted with high probabilities, and clinical features of indications. Second, the progressive report decoder consists of a sentence decoder and a word decoder, where we propose image-sentence matching and description accuracy losses to constrain the visual-textual semantic consistency. Extensive experiments on the public MIMIC-CXR and IU X-Ray datasets show that our model achieves consistent improvements over the state-of-the-art methods.

A recent trend in computer vision is to replace convolutions with transformers. However, the performance gain of transformers is attained at a steep cost, requiring GPU years and hundreds of millions of samples for training. This excessive resource usage compensates for a misuse of transformers: Transformers densely model relationships between its inputs -- ideal for late stages of a neural network, when concepts are sparse and spatially-distant, but extremely inefficient for early stages of a network, when patterns are redundant and localized. To address these issues, we leverage the respective strengths of both operations, building convolution-transformer hybrids. Critically, in sharp contrast to pixel-space transformers, our Visual Transformer (VT) operates in a semantic token space, judiciously attending to different image parts based on context. Our VTs significantly outperforms baselines: On ImageNet, our VT-ResNets outperform convolution-only ResNet by 4.6 to 7 points and transformer-only ViT-B by 2.6 points with 2.5 times fewer FLOPs, 2.1 times fewer parameters. For semantic segmentation on LIP and COCO-stuff, VT-based feature pyramid networks (FPN) achieve 0.35 points higher mIoU while reducing the FPN module's FLOPs by 6.5x.

State-of-the-art approaches for visually-guided audio source separation typically assume sources that have characteristic sounds, such as musical instruments. These approaches often ignore the visual context of these sound sources or avoid modeling object interactions that may be useful to better characterize the sources, especially when the same object class may produce varied sounds from distinct interactions. To address this challenging problem, we propose Audio Visual Scene Graph Segmenter (AVSGS), a novel deep learning model that embeds the visual structure of the scene as a graph and segments this graph into subgraphs, each subgraph being associated with a unique sound obtained by co-segmenting the audio spectrogram. At its core, AVSGS uses a recursive neural network that emits mutually-orthogonal sub-graph embeddings of the visual graph using multi-head attention. These embeddings are used for conditioning an audio encoder-decoder towards source separation. Our pipeline is trained end-to-end via a self-supervised task consisting of separating audio sources using the visual graph from artificially mixed sounds. In this paper, we also introduce an ""in the wild" video dataset for sound source separation that contains multiple non-musical sources, which we call Audio Separation in the Wild (ASIW). This dataset is adapted from the AudioCaps dataset, and provides a challenging, natural, and daily-life setting for source separation. Thorough experiments on the proposed ASIW and the standard MUSIC datasets demonstrate state-of-the-art sound separation performance of our method against recent prior approaches.

Existing state-of-the-art saliency detection methods heavily rely on CNN-based architectures. Alternatively, we rethink this task from a convolution-free sequence-to-sequence perspective and predict saliency by modeling long-range dependencies, which can not be achieved by convolution. Specifically, we develop a novel unified model based on a pure transformer, namely, Visual Saliency Transformer (VST), for both RGB and RGB-D salient object detection (SOD). It takes image patches as inputs and leverages the transformer to propagate global contexts among image patches. Unlike conventional architectures used in Vision Transformer (ViT), we leverage multi-level token fusion and propose a new token upsampling method under the transformer framework to get high-resolution detection results. We also develop a token-based multi-task decoder to simultaneously perform saliency and boundary detection by introducing task-related tokens and a novel patch-task-attention mechanism. Experimental results show that our model outperforms existing methods on both RGB and RGB-D SOD benchmark datasets. Most importantly, our whole framework not only provides a new perspective for the SOD field but also shows a new paradigm for transformer-based dense prediction models. Code is available at https://github.com/nnizhang/VST.

Computer vision applications such as visual relationship detection and human object interaction can be formulated as a composite (structured) set detection problem in which both the parts (subject, object, and predicate) and the sum (triplet as a whole) are to be detected in a hierarchical fashion. In this paper, we present a new approach, denoted Part-and-Sum detection Transformer (PST), to perform end-to-end visual composite set detection. Different from existing Transformers in which queries are at a single level, we simultaneously model the joint part and sum hypotheses/interactions with composite queries and attention modules. We explicitly incorporate sum queries to enable better modeling of the part-and-sum relations that are absent in the standard Transformers. Our approach also uses novel tensor-based part queries and vector-based sum queries, and models their joint interaction. We report experiments on two vision tasks, visual relationship detection and human object interaction and demonstrate that PST achieves state of the art results among single-stage models, while nearly matching the results of custom designed two-stage models.

We present a novel graph-structured memory for visual navigation, called visual graph memory (VGM), which consists of unsupervised image representations obtained from navigation history. The proposed VGM is constructed incrementally based on the similarities among the unsupervised representations of observed images, and these representations are learned from an unlabeled image dataset. We also propose a navigation framework that can utilize the proposed VGM to tackle visual navigation problems. By incorporating a graph convolutional network and the attention mechanism, the proposed agent refers to the VGM to navigate the environment while simultaneously building the VGM. Using the VGM, the agent can embed its navigation history and other useful task-related information. We validate our approach on the visual navigation tasks using the Habitat simulator with the Gibson dataset, which provides a photo-realistic simulation environment. The extensive experimental results show that the proposed navigation agent with VGM surpasses the state-of-the-art approaches on image-goal navigation tasks.

Scene graph generation aims to identify objects and their relations in images, providing structured image representations that can facilitate numerous applications in computer vision. However, scene graph models usually require supervised learning on large quantities of labeled data with intensive human annotation. In this work, we propose visual distant supervision, a novel paradigm of visual relation learning, which can train scene graph models without any human-labeled data. The intuition is that by aligning commonsense knowledge bases and images, we can automatically create large-scale labeled data to provide distant supervision for visual relation learning. To alleviate the noise in distantly labeled data, we further propose a framework that iteratively estimates the probabilistic relation labels and eliminates the noisy ones. Comprehensive experimental results show that our distantly supervised model outperforms strong weakly supervised and semi-supervised baselines. By further incorporating human-labeled data in a semi-supervised fashion, our model outperforms state-of-the-art fully supervised models by a large margin (e.g., 8.3 micro- and 7.8 macro-recall@50 improvements for predicate classification in Visual Genome evaluation). We make the data and code for this paper publicly available at https://github.com/thunlp/VisualDS.

Vision-based Continuous Sign Language Recognition (CSLR) aims to recognize unsegmented signs from image streams. Overfitting is one of the most critical problems in CSLR training, and previous works show that the iterative training scheme can partially solve this problem while also costing more training time. In this study, we revisit the iterative training scheme in recent CSLR works and realize that sufficient training of the feature extractor is critical to solving the overfitting problem. Therefore, we propose a Visual Alignment Constraint (VAC) to enhance the feature extractor with alignment supervision. Specifically, the proposed VAC comprises two auxiliary losses: one focuses on visual features only, and the other enforces prediction alignment between the feature extractor and the alignment module. Moreover, we propose two metrics to reflect overfitting by measuring the prediction inconsistency between the feature extractor and the alignment module. Experimental results on two challenging CSLR datasets show that the proposed VAC makes CSLR networks end-to-end trainable and achieves competitive performance.

In this work, we address the challenging task of referring segmentation. The query expression in referring segmentation typically indicates the target object by describing its relationship with others. Therefore, to find the target one among all instances in the image, the model must have a holistic understanding of the whole image. To achieve this, we reformulate referring segmentation as a direct attention problem: finding the region in the image where the query language expression is most attended to. We introduce transformer and multi-head attention to build a network with an encoder-decoder attention mechanism architecture that "queries" the given image with the language expression. Furthermore, we propose a Query Generation Module, which produces multiple sets of queries with different attention weights that represent the diversified comprehensions of the language expression from different aspects. At the same time, to find the best way from these diversified comprehensions based on visual clues, we further propose a Query Balance Module to adaptively select the output features of these queries for a better mask generation. Without bells and whistles, our approach is light-weight and achieves new state-of-the-art performance consistently on three referring segmentation datasets, RefCOCO, RefCOCO+, and G-Ref. Our code is available at https://github.com/henghuiding/Vision-Language-Transformer.

Vision-language Navigation (VLN) task requires an agent to perceive both the visual scene and natural language and navigate step-by-step. Large data bias makes the VLN task challenging, which is caused by the disparity ratio between small data scale and large navigation space. Previous works have proposed many data augmentation methods to reduce data bias. However, these works do not explicitly reduce the data bias across different house scenes. Therefore, the agent would be overfitting to the seen scenes and perform navigation poorly in the unseen scenes. To tackle this problem, we propose the random environmental mixup (REM) method, which generates augmentation data in cross-connected house scenes. This method consists of three steps: 1) we select the key viewpoints according to the room connection graph for each scene in the training split; 2) we cross-connect the key views of different scenes to construct augmented scenes; 3) we generate augmentation data triplets (environment, path, instruction) in the cross-connected scenes. Our experiments prove that the augmentation data helps the agent reduce its performance gap between the seen and unseen environment and improve its performance, making our model be the best existing approach on the standard benchmark.