ICCV 2021 Papers

The goal of this paper is to self-train a 3D convolutional neural network on an unlabeled video collection for deployment on small-scale video collections. As smaller video datasets benefit more from motion than appearance, we strive to train our network using optical flow, but avoid its computation during inference. We propose the first motion-augmented self-training regime, we call MotionFit. We start with supervised training of a motion model on a small, and labeled, video collection. With the motion model we generate pseudo-labels for a large unlabeled video collection, which enables us to transfer knowledge by learning to predict these pseudo-labels with an appearance model. Moreover, we introduce a multi-clip loss as a simple yet efficient way to improve the quality of the pseudo-labeling, even without additional auxiliary tasks. We also take into consideration the temporal granularity of videos during self-training of the appearance model, which was missed in previous works. As a result we obtain a strong motion-augmented representation model suited for video downstream tasks like action recognition and clip retrieval. On small-scale video datasets, MotionFit outperforms alternatives for knowledge transfer by 5%-8%, video-only self-supervision by 1%-7% and semisupervised learning by 9%-18% using the same amount of class labels.

Predicting human motion from a historical pose sequence is at the core of many applications in computer vision. Current state-of-the-art methods concentrate on learning motion contexts in the pose space, however, the high dimensionality and complex nature of human pose invoke inherent difficulties in extracting such contexts. In this paper, we instead advocate to model motion contexts in the joint trajectory space, as the trajectory of a joint is smooth, vectorial, and gives sufficient information to the model. Moreover, most existing methods consider only the dependencies between skeletal connected joints, disregarding prior knowledge and the hidden connections between geometrically separated joints. Motivated by this, we present a semi-constrained graph to explicitly encode skeletal connections and prior knowledge, while adaptively learn implicit dependencies between joints. We also explore the applications of our approach to a range of objects including human, fish, and mouse. Surprisingly, our method sets the new state-of-the-art performance on 4 different benchmark datasets, a remarkable highlight is that it achieves a 19.1% accuracy improvement over current state-of-the-art in average. To facilitate future research, we have released our code at https://github.com/Pose-Group/MPT.

Video captioning is an important vision task and has been intensively studied in the computer vision community. Existing methods that utilize the fine-grained spatial information have achieved significant improvements, however, they either rely on costly external object detectors or do not sufficiently model the spatial/temporal relations. In this paper, we aim at designing a spatial information extraction and aggregation method for video captioning without the need of external object detectors. For this purpose, we propose a Recurrent Region Attention module to better extract diverse spatial features, and by employing Motion-Guided Cross-frame Message Passing, our model is aware of the temporal structure and able to establish high-order relations among the diverse regions across frames. They jointly encourage information communication and produce compact and powerful video representations. Furthermore, an Adjusted Temporal Graph Decoder is proposed to flexibly update video features and model high-order temporal relations during decoding. Experimental results on three benchmark datasets: MSVD, MSR-VTT, and VATEX demonstrate that our proposed method can outperform state-of-the-art methods.

We present a dual-pathway approach for recognizing fine-grained interactions from videos. We build on the success of prior dual-stream approaches, but make a distinction between the static and dynamic representations of objects and their interactions explicit by introducing separate motion and object detection pathways. Then, using our new Motion-Guided Attention Fusion module, we fuse the bottom-up features in the motion pathway with features captured from object detections to learn the temporal aspects of an action. We show that our approach can generalize across appearance effectively and recognize actions where an actor interacts with previously unseen objects. We validate our approach using the compositional action recognition task from the Something-Something-v2 dataset where we outperform existing state-of-the-art methods. We also show that our method can generalize well to real world tasks by showing state-of-the-art performance on recognizing humans assembling various IKEA furniture on the IKEA-ASM dataset.

In this paper, we propose an end-to-end learning framework for event-based motion deblurring in a self-supervised manner, where real-world events are exploited to alleviate the performance degradation caused by data inconsistency. To achieve this end, optical flows are predicted from events, with which the blurry consistency and photometric consistency are exploited to enable self-supervision on the deblurring network with real-world data. Furthermore, a piece-wise linear motion model is proposed to take into account motion non-linearities and thus leads to an accurate model for the physical formation of motion blurs in the real-world scenario. Extensive evaluation on both synthetic and real motion blur datasets demonstrates that the proposed algorithm bridges the gap between simulated and real-world motion blurs and shows remarkable performance for event-based motion deblurring in real-world scenarios.

In this paper, we introduce a new framework for unsupervised deep homography estimation. Our contributions are 3 folds. First, unlike previous methods that regress 4 offsets for a homography, we propose a homography flow representation, which can be estimated by a weighted sum of 8 pre-defined homography flow bases. Second, considering a homography contains 8 Degree-of-Freedoms (DOFs) that is much less than the rank of the network features, we propose a Low Rank Representation (LRR) block that reduces the feature rank, so that features corresponding to the dominant motions are retained while others are rejected. Last, we propose a Feature Identity Loss (FIL) to enforce the learned image feature warp-equivariant, meaning that the result should be identical if the order of warp operation and feature extraction is swapped. With this constraint, the unsupervised optimization is achieved more effectively and more stable features are learned. Extensive experiments are conducted to demonstrate the effectiveness of all the newly proposed components, and results show that our approach outperforms the state-of-the-art on the homography benchmark datasets both qualitatively and quantitatively. Code is available at https://github.com/megvii-research/BasesHomo

Human pose estimation from videos has many real-world applications. Existing methods focus on applying models with a uniform computation profile on fully de- coded frames, ignoring the freely available motion signals and motion-compensation residuals from the compressed stream. A novel model, called Motion Adaptive Pose Net is proposed to exploit the compressed streams to efficiently decode pose sequences from videos. The model incorporates a Motion Compensated ConvLSTM to propagate the spatially aligned features, along with an adaptive gate to dynamically determine if the computationally expensive features should be extracted from fully decoded frames to compensate the motion-warped features, solely based on the residual errors. Leveraging the informative yet readily available signals from compressed streams, we propagate the latent features through our Motion Adaptive Pose Net efficiently. Our model outperforms the state-of-the-art models in pose- estimation accuracy on two widely used datasets with only around half of the computation complexity.

Many objects do not appear frequently enough in complex scenes (e.g., certain handbags in living rooms) for training an accurate object detector, but are often found frequently by themselves (e.g., in product images). Yet, these object-centric images are not effectively leveraged for improving object detection in scene-centric images. In this paper, we propose Mosaic of Object-centric images as Scene-centric images (MosaicOS), a simple and novel framework that is surprisingly effective at tackling the challenges of long-tailed object detection. Keys to our approach are three-fold: (i) pseudo scene-centric image construction from object-centric images for mitigating domain differences, (ii) high-quality bounding box imputation using the object-centric images' class labels, and (iii) a multi-stage training procedure. On LVIS object detection (and instance segmentation), MosaicOS leads to a massive 60% (and 23%) relative improvement in average precision for rare object categories. We also show that our framework can be compatibly used with other existing approaches to achieve even further gains. Our pre-trained models are publicly available at https://github.com/czhang0528/MosaicOS/.

Many emerging applications of intelligent robots need to explore and understand new environments, where it is desirable to detect objects of novel categories on the fly with minimum online efforts. This is an object detection on demand (ODOD) task. It is challenging, because it is impossible to annotate large data on the fly, and the embedded systems are usually unable to perform back-propagation which is essential for training. Most existing few-shot detection methods are confronted here as they need extra training. We propose a novel morphable detector (MD), that simply "morphs"" some of its changeable parameters online estimated from the few samples, so as to detect novel categories without any extra training. The MD has two sets of parameters, one for the feature embedding and the other for category representation(called "prototypes""). Each category is associated with a hidden prototype to be learned by integrating the visual and semantic embeddings. The learning of the MD is based on the alternate learning of the feature embedding and the prototypes in an EM-like approach which allows the recovery of an unknown prototype from a few samples of a novel category. Once an MD is learned, it is able to use a few samples of a novel category to directly compute its prototype to fulfill the online morphing process. We have shown the superiority of the MD in Pascal, COCO and FSOD datasets.

We propose a novel method for reconstructing floor plans from noisy 3D point clouds. Our main contribution is a principled approach that relies on the Monte Carlo Tree Search (MCTS) algorithm to maximize a suitable objective function efficiently despite the complexity of the problem. Like previous work, we first project the input point cloud to a top view to create a density map and extract room proposals from it. Our method selects and optimizes the polygonal shapes of these room proposals jointly to fit the density map and outputs an accurate vectorized floor map even for large complex scenes. To do this, we adapted MCTS, an algorithm originally designed to learn to play games, to select the room proposals by maximizing an objective function combining the fitness with the density map as predicted by a deep network and regularizing terms on the room shapes. We also introduce a refinement step to MCTS that adjusts the shape of the room proposals. For this step, we propose a novel differentiable method for rendering the polygonal shapes of these proposals. We evaluate our method on the recent and challenging Structured3D and Floor-SP datasets and show a significant improvement over the state-of-the-art, without imposing any hard constraints nor assumptions on the floor plan configurations.

This paper focuses on the regression of multiple 3D people from a single RGB image. Existing approaches predominantly follow a multi-stage pipeline that first detects people in bounding boxes and then independently regresses their 3D body meshes. In contrast, we propose to Regress all meshes in a One-stage fashion for Multiple 3D People (termed ROMP). The approach is conceptually simple, bounding box-free, and able to learn a per-pixel representation in an end-to-end manner. Our method simultaneously predicts a Body Center heatmap and a Mesh Parameter map, which can jointly describe the 3D body mesh on the pixel level. Through a body-center-guided sampling process, the body mesh parameters of all people in the image are easily extracted from the Mesh Parameter map. Equipped with such a fine-grained representation, our one-stage framework is free of the complex multi-stage process and more robust to occlusion. Compared with state-of-the-art methods, ROMP achieves superior performance on the challenging multi-person benchmarks, including 3DPW and CMU Panoptic. Experiments on crowded/occluded datasets demonstrate the robustness under various types of occlusion. The code, released at https://github.com/Arthur151/ROMP, is the first real-time implementation of monocular multi-person 3D mesh regression.

Self-supervised depth estimation for indoor environments is more challenging than its outdoor counterpart in at least the following two aspects: (i) the depth range of indoor sequences varies a lot across different frames, making it difficult for the depth network to induce consistent depth cues, whereas the maximum distance in outdoor scenes mostly stays the same as the camera usually sees the sky; (ii) the indoor sequences contain much more rotational motions, which cause difficulties for the pose network, while the motions of outdoor sequences are pre-dominantly translational, especially for driving datasets such as KITTI. In this paper, special considerations are given to those challenges and a set of good practices are consolidated for improving the performance of self-supervised monocular depth estimation in indoor environments. The proposed method mainly consists of two novel modules, i.e., a depth factorization module and a residual pose estimation module, each of which is designed to respectively tackle the aforementioned challenges. The effectiveness of each module is shown through a carefully conducted ablation study and the demonstration of the state-of-the-art performance on three indoor datasets, i.e., EuRoC, NYUv2 and 7-Scenes.

Multi-Layer Perceptrons (MLPs) make powerful functional representations for sampling and reconstruction problems involving low-dimensional signals like images,shapes and light fields. Recent works have significantly improved their ability to represent high-frequency content by using periodic activations or positional encodings. This often came at the expense of generalization: modern methods are typically optimized for a single signal. We present a new representation that generalizes to multiple instances and achieves state-of-the-art fidelity. We use a dual-MLP architecture to encode the signals. A synthesis network creates a functional mapping from a low-dimensional input(e.g. pixel-position) to the output domain (e.g. RGB color).A modulation network maps a latent code corresponding to the target signal to parameters that modulate the periodic activations of the synthesis network. We also propose a local-functional representation which enables generalization. The signal's domain is partitioned into a regular grid,with each tile represented by a latent code. At test time, the signal is encoded with high-fidelity by inferring (or directly optimizing) the latent code-book. Our approach produces generalizable functional representations of images, videos and shapes, and achieves higher reconstruction quality than prior works that are optimized for a single signal.

The graph convolutional network (GCN) has recently achieved promising performance of 3D human pose estimation (HPE) by modeling the relationship among body parts. However, most prior GCN approaches suffer from two main drawbacks. First, they share a feature transformation for each node within a graph convolution layer. This prevents them from learning different relations between different body joints. Second, the graph is usually defined according to the human skeleton and is suboptimal because human activities often exhibit motion patterns beyond the natural connections of body joints. To address these limitations, we introduce a novel Modulated GCN for 3D HPE. It consists of two main components: weight modulation and affinity modulation. Weight modulation learns different modulation vectors for different nodes so that the feature transformations of different nodes are disentangled while retaining a small model size. Affinity modulation adjusts the graph structure in a GCN so that it can model additional edges beyond the human skeleton. We investigate several affinity modulation methods as well as the impact of regularizations. Rigorous ablation study indicates both types of modulation improve performance with negligible overhead. Compared with state-of-the-art GCNs for 3D HPE, our approach either significantly reduces the estimation errors, e.g., by around 10%, while retaining a small model size or drastically reduces the model size, e.g., from 4.22M to 0.29M (a 14.5 X reduction), while achieving comparable performance. Results on two benchmarks show our Modulated GCN outperforms some recent states of the art. Our code is available at https://github.com/ZhimingZo/Modulated-GCN.

This paper presents a self-supervised method for learning reliable visual correspondence from unlabeled videos. We formulate the correspondence as finding paths in a joint space-time graph, where nodes are grid patches sampled from frames, and are linked by two type of edges: (i) neighbor relations that determine the aggregation strength from intra-frame neighbors in space, and (ii) similarity relations that indicate the transition probability of inter-frame paths across time. Leveraging the cycle-consistency in videos, our contrastive learning objective discriminates dynamic objects from both their neighboring views and temporal views. Compared with prior works, our approach actively explores the neighbor relations of central instances to learn a latent association between center-neighbor pairs (eg, "hand -- arm") across time, thus improving the instance discrimination. Without fine-tuning, our learned representation outperforms the state-of-the-art self-supervised methods on a variety of visual tasks including video object propagation, part propagation, and pose keypoint tracking. Our self-supervised method also surpasses some fully supervised algorithms designed for the specific tasks.

We introduce Mixture-based Feature Space Learning (MixtFSL) for obtaining a rich and robust feature representation in the context of few-shot image classification. Previous works have proposed to model each base class either with a single point or with a mixture model by relying on offline clustering algorithms. In contrast, we propose to model base classes with mixture models by simultaneously training the feature extractor and learning the mixture model parameters in an online manner. This results in a richer and more discriminative feature space which can be employed to classify novel examples from very few samples. Two main stages are proposed to train the MixtFSL model. First, the multimodal mixtures for each base class and the feature extractor parameters are learned using a combination of two loss functions. Second, the resulting network and mixture models are progressively refined through a leader-follower learning procedure, which uses the current estimate as a "target" network. This target network is used to make a consistent assignment of instances to mixture components, which increases performance and stabilizes training. The effectiveness of our end-to-end feature space learning approach is demonstrated with extensive experiments on four standard datasets and four backbones. Notably, we demonstrate that when we combine our robust representation with recent alignment-based approaches, we achieve new state-of-the-art results in the inductive setting, with an absolute accuracy for 5-shot classification of 82.45 on miniImageNet, 88.20 with tieredImageNet, and 60.70 in FC100 using the ResNet-12 backbone.

It is common practice to represent spoken languages at their phonetic level. However, for sign languages, this implies breaking motion into its constituent motion primitives. Avatar based Sign Language Production (SLP) has traditionally done just this, building up animation from sequences of hand motions, shapes and facial expressions. However, more recent deep learning based solutions to SLP have tackled the problem using a single network that estimates the full skeletal structure. We propose splitting the SLP task into two distinct jointly-trained sub-tasks. The first translation sub-task translates from spoken language to a latent sign language representation, with gloss supervision. Subsequently, the animation sub-task aims to produce expressive sign language sequences that closely resemble the learnt spatio-temporal representation. Using a progressive transformer for the translation sub-task, we propose a novel Mixture of Motion Primitives (MoMP) architecture for sign language animation. A set of distinct motion primitives are learnt during training, that can be temporally combined at inference to animate continuous sign language sequences. We evaluate on the challenging RWTH-PHOENIX-Weather-2014T(PHOENIX14T) dataset, presenting extensive ablation studies and showing that MoMP outperforms baselines in user evaluations. We achieve state-of-the-art back translation performance with an 11% improvement over competing results. Importantly, and for the first time, we showcase stronger performance for a full translation pipeline going from spoken language to sign, than from gloss to sign.

Recent strategies achieved ensembling ""for free"" by fitting concurrently diverse subnetworks inside a single base network. The main idea during training is that each subnetwork learns to classify only one of the multiple inputs simultaneously provided. However, the question of how to best mix these multiple inputs has not been studied so far. In this paper, we introduce MixMo, a new generalized framework for learning multi-input multi-output deep subnetworks. Our key motivation is to replace the suboptimal summing operation hidden in previous approaches by a more appropriate mixing mechanism. For that purpose, we draw inspiration from successful mixed sample data augmentations. We show that binary mixing in features - particularly with rectangular patches from CutMix - enhances results by making subnetworks stronger and more diverse. We improve state of the art for image classification on CIFAR-100 and Tiny ImageNet datasets. Our easy to implement models notably outperform data augmented deep ensembles, without the inference and memory overheads. As we operate in features and simply better leverage the expressiveness of large networks, we open a new line of research complementary to previous works.

User data confidentiality protection is becoming a rising challenge in the present deep learning research. Without access to data, conventional data-driven model compression faces a higher risk of performance degradation. Recently, some works propose to generate images from a specific pretrained model to serve as training data. However, the inversion process only utilizes biased feature statistics stored in one model and is from low-dimension to high-dimension. As a consequence, it inevitably encounters the difficulties of generalizability and inexact inversion, which leads to unsatisfactory performance. To address these problems, we propose MixMix based on two simple yet effective techniques: (1) Feature Mixing: utilizes various models to construct a universal feature space for generalized inversion; (2) Data Mixing: mixes the synthesized images and labels to generate exact label information. We prove the effectiveness of MixMix from both theoretical and empirical perspectives. Extensive experiments show that MixMix outperforms existing methods on the mainstream compression tasks, including quantization, knowledge distillation and pruning. Specifically, MixMix achieves up to 4% and 20% accuracy uplift on quantization and pruning, respectively, compared to existing data-free compression work.