ECCV 2020 Papers

Despite the great progress in human motion prediction, it remains a challenging task due to the complicated structural dynamics of human behaviors. In this paper, we address this problem in three aspects. First, to capture the long-range spatial correlations and temporal dependencies, we apply a transformer-based architecture with the global attention mechanism. Specifically, we feed the network with the sequential joints encoded with the temporal information for spatial and temporal explorations. Second, to further exploit the inherent kinematic chains for better 3D structures, we apply a progressive-decoding strategy, which performs in a central-to-peripheral extension according to the structural connectivity. Last, in order to incorporate a general motion space for high-quality prediction, we build a memory-based dictionary, which aims to preserve the global motion patterns in training data to guide the predictions. We evaluate the proposed method on two challenging benchmark datasets (Human3.6M and CMU-Mocap). Experimental results show our superior performance compared with the state-of-the-art approaches."

Learning predictive models from interaction with the world allows an agent, such as a robot, to learn about how the world works, and then use this learned model to plan coordinated sequences of actions to bring about desired outcomes. However, learning a model that captures the dynamics of complex skills represents a major challenge: if the agent needs a good model to perform these skills, it might never be able to collect the experience on its own that is required to learn these delicate and complex behaviors. Instead, we can imagine augmenting the training set with observational data of other agents, such as humans. Such data is likely more plentiful, but cannot always be combined with data from the original agent. For example, videos of humans might show a robot how to use a tool, but (i) are not annotated with suitable robot actions, and (ii) contain a systematic distributional shift due to the embodiment differences between humans and robots. We address the first challenge by formulating the corresponding graphical model and treating the action as an observed variable for the interaction data and an unobserved variable for the observation data, and the second challenge by using a domain-dependent prior. In addition to interaction data, our method is able to leverage videos of passive observations in a driving dataset and a dataset of robotic manipulation videos to improve video prediction performance. We also demonstrate that, in a real-world tabletop robotic manipulation setting, we are able to significantly improve control performance by learning a model from both robot data and observations of humans."

Many real-world tasks such as classification of digital histopathological images and 3D object detection involve learning from a set of instances. In these cases, only a group of instances or a set, collectively, contains meaningful information and therefore only the sets have labels, and not individual data instances. In this work, we present a permutation invariant neural network called Memory-based Exchangeable Model (MEM) for learning universal set functions. The MEM model consists of memory units that embed an input sequence to high-level features enabling it to learn inter-dependencies among instances through a self-attention mechanism. We evaluated the learning ability of MEM on various toy datasets, point cloud classification, and classification of whole slide images (WSIs) into two subtypes of the lung cancer--Lung Adenocarcinoma, and Lung Squamous Cell Carcinoma. We systematically extracted patches from WSIs of the lung, downloaded from The Cancer Genome Atlas (TCGA) dataset, the largest public repository of WSIs, achieving a competitive accuracy of 84.84% for classification of two sub-types of lung cancer. The results on other datasets are promising as well, and demonstrate the efficacy of our model. \keywords{Permutation Invariant Models, Multi Instance Learning, Whole Slide Image Classification, Medical Images}"

This paper proposes a novel single-image piecewise planar reconstruction technique that infers and enforces inter-plane relationships. Our approach takes a planar reconstruction result from an existing system, then utilizes convolutional neural network (CNN) to (1) classify if two planes are orthogonal or parallel; and 2) infer if two planes are touching and, if so, where in the image. We formulate an optimization problem to refine plane parameters and employ a message passing neural network to refine plane segmentation masks by enforcing the inter-plane relations. Our qualitative and quantitative evaluations demonstrate the effectiveness of the proposed approach in terms of plane parameters and segmentation accuracy."

Open set recognition is an emerging research area that aims to simultaneously classify samples from predefined classes and identify the rest as 'unknown'. In this process, one of the key challenges is to reduce the risk of generalizing the inherent characteristics of numerous unknown samples learned from a small amount of known data. In this paper, we propose a new concept, Reciprocal Point, which is the potential representation of the extra-class space corresponding to each known category. The sample can be classified to known or unknown by the otherness with reciprocal points. To tackle the open set problem, we offer a novel open space risk regularization term. Based on the bounded space constructed by reciprocal points, the risk of unknown is reduced through multi-category interaction. The novel learning framework called Reciprocal Point Learning (RPL), which can indirectly introduce the unknown information into the learner with only known classes, so as to learn more compact and discriminative representations. Moreover, we further construct a new large-scale challenging aircraft dataset for open set recognition: Aircraft 300 (Air-300). Extensive experiments on multiple benchmark datasets indicate that our framework is significantly superior to other existing approaches and achieves state-of-the-art performance on standard open set benchmarks."

Target-driven visual navigation aims at navigating an agent towards a given target based on the observation of the agent. In this task, it is critical to learn informative visual representation and robust navigation policy. Aiming to improve these two components, this paper proposes three complementary techniques, object relation graph (ORG), trial-driven imitation learning (IL), and a memory-augmented tentative policy network (TPN). ORG improves visual representation learning by integrating object relationships, including category closeness and spatial correlations, mph{e.g.,} a TV usually co-occurs with a remote spatially. Both Trial-driven IL and TPN underlie robust navigation policy, instructing the agent to escape from deadlock states, such as looping or being stuck. Specifically, trial-driven IL is a type of supervision used in policy network training, while TPN, mimicking the IL supervision in unseen environment, is applied in testing. %instructing an agent to escape from deadlock states. Experiment in the artificial environment AI2-Thor validates that each of the techniques is effective. When combined, the techniques bring significantly improvement over baseline methods in navigation effectiveness and efficiency in unseen environments. We report 22.8\% and 23.5\% increase in success rate and Success weighted by Path Length (SPL), respectively. The code is available at \url{https://github.com/xiaobaishu0097/ECCV-VN.git}."

We study the problem of common sense placement of visual objects in an image.  This involves multiple aspects of visual recognition: the instance segmentation of the scene, 3D layout, and common knowledge of how objects are placed and where objects are moving in the 3D scene. This seemingly simple task is difficult for current learning-based approaches because of the lack of labeled training of foreground objects paired with cleaned background scenes. We propose a self-learning framework that automatically generates the necessary training data without any manual labeling by detecting, cutting, and inpainting objects from an image.  We propose a PlaceNet that predicts a diverse distribution of common sense locations when given a foreground object and a background scene. We show one practical use of our object placement network for augmenting training datasets by recomposition of object-scene with a key property of contextual relationship preservation. We demonstrate improvement of object detection and instance segmentation performance on both Cityscape and datasets.  We also show that the learned representation of our PlaceNet displays strong discriminative power in image retrieval and classification."

Object Permanence allows people to reason about the location of non-visible objects, by understanding that they continue to exist even when not perceived directly. Object Permanence is critical for building a model of the world, since objects in natural visual scenes dynamically occlude and contain each-other. Intensive studies in developmental psychology, suggesting that object permanence is a challenging task that is learned through extensive experience.

Video object segmentation, i.e., the separation of a target object from background in video, has made significant progress on real and challenging videos in recent years. To leverage this progress in 3D applications, this paper addresses the problem of learning to estimate the depth of segmented objects given some measurement of camera motion (e.g., from robot kinematics or vehicle odometry). We achieve this by, first, introducing a diverse, extensible dataset and, second, designing a novel deep network that estimates the depth of objects using only segmentation masks and uncalibrated camera movement. Our data-generation framework creates artificial object segmentations that are scaled for changes in distance between the camera and object, and our network learns to estimate object depth even with segmentation errors. We demonstrate our approach across domains using a robot camera to locate objects from the YCB dataset and a vehicle camera to locate obstacles while driving."

In this paper, we propose a noise-aware encoder-decoder framework to disentangle a clean saliency predictor from noisy training examples, where the noisy labels are generated by unsupervised handcrafted feature-based methods. The proposed model consists of two sub-models parameterized by neural networks: (1) a saliency predictor that maps input images to clean saliency maps, and (2) a noise generator, which is a latent variable model that produces noises from Gaussian latent vectors. The whole model that represents noisy labels is a sum of the two sub-models. The goal of training the model is to estimate the parameters of both sub-models, and simultaneously infer the corresponding latent vector of each noisy label. We propose to train the model by using an alternating back-propagation (ABP) algorithm, which alternates the following two steps: (1) learning back-propagation for estimating the parameters of two sub-models by gradient ascent, and (2) inferential back-propagation for inferring the latent vectors of training noisy examples by Langevin Dynamics. To prevent the network from converging to trivial solutions, we utilize an edge-aware smoothness loss to regularize hidden saliency maps to have similar structures as their corresponding images. Experimental results on several benchmark datasets indicate the effectiveness of the proposed model. "

This paper studies the fundamental problem of learning deep generative models that consist of multiple layers of latent variables organized in top-down architectures. Such models have high expressivity and allow for learning hierarchical representations. Learning such a generative model requires inferring the latent variables for each training example based on the posterior distribution of these latent variables. The inference typically requires Markov chain Monte Caro (MCMC) that can be time consuming. In this paper, we propose to use noise initialized non-persistent short run MCMC, such as finite step Langevin dynamics initialized from the prior distribution of the latent variables, as an approximate inference engine, where the step size of the Langevin dynamics is variationally optimized by minimizing the Kullback-Leibler divergence between the distribution produced by the short run MCMC and the posterior distribution. Our experiments show that the proposed method outperforms variational auto-encoder (VAE) in terms of reconstruction error and synthesis quality. The advantage of the proposed method is that it is simple and automatic without the need to design an inference model."

Monocular visual odometry (VO) suffers severely from error accumulation during frame-to-frame pose estimation. In this paper, we present a self-supervised learning method for VO with special consideration for consistency over longer sequences. To this end, we model the long-term dependency in pose prediction using a pose network that features a two-layer convolutional LSTM module. We train the networks with purely self-supervised losses, including a cycle consistency loss that mimics the loop closure module in geometric VO. Inspired by prior geometric systems, we allow the networks to see beyond a small temporal window during training, through a novel a loss that incorporates temporally distant (g $O(100)$) frames. Given GPU memory constraints, we propose a stage-wise training mechanism, where the first stage operates in a local time window and the second stage refines the poses with a ``global'' loss given the first stage features. We demonstrate competitive results on several standard VO datasets, including KITTI and TUM RGB-D."

Automatically generating sentences to describe events and temporally localizing sentences in a video are two important tasks that bridge language and videos. Recent techniques leverage the multimodal nature of videos by using off-the-shelf features to represent videos, but interactions between modalities are rarely explored. Inspired by the fact that there exist cross-modal interactions in the human brain, we propose a novel method for learning pairwise modality interactions in order to better exploit complementary information for each pair of modalities in videos and thus improve performances on both tasks. We model modality interaction in both the sequence and channel levels in a pairwise fashion, and the pairwise interaction also provides some explainability for the predictions of target tasks. We demonstrate the effectiveness of our method and validate specific design choices through extensive ablation studies. Our method turns out to achieve state-of-the-art performances on four standard benchmark datasets: MSVD and MSR-VTT (event captioning task), and Charades-STA and ActivityNet Captions (temporal sentence localization task)."

In this paper, we propose a cascading network for compressed sensing of images with progressive reconstruction. Specifically, we decompose the complex reconstruction mapping into the cascade of incremental detail reconstruction (IDR) modules and measurement residual updating (MRU) modules. The IDR module is designed to reconstruct the remaining details from the residual measurement vector, and MRU is employed to update the residual measurement vector and feed it into the next IDR module. The contextual memory module is introduced to augment the capacity of IDR modules, therefore facilitating the information interaction among all the IDR modules. The final reconstruction is calculated by accumulating the outputs of all the IDR modules. Extensive experiments on natural images and magnetic resonance images demonstrate the proposed method achieves better performance against the state-of-the-art methods."

We propose a motion forecasting model that exploits a novel structured map representation as well as actor-map interactions. Instead of encoding vectorized maps as raster images, we construct a lane graph from raw map data to explicitly preserve the map structure. To capture the complex topology and long range dependencies of the lane graph, we propose LaneGCN which extends graph convolutions with multiple adjacency matrices and along-lane dilation. To capture the complex interactions between actors and maps, we exploit a fusion network consisting of four types of interactions, actor-to-lane, lane-to-lane, lane-to-actor and actor-to-actor. Powered by LaneGCN and actor-map interactions, our model is able to predict accurate and realistic multi-modal trajectories. Our approach significantly outperforms the state-of-the-art on the large scale Argoverse motion forecasting benchmark."

Interactive image retrieval is an emerging research topic with the objective of integrating inputs from multiple modalities as query for retrieval, e.g., textual feedback from users to guide, modify or refine image retrieval. In this work, we study the problem of composing images and textual modifications for language-guided retrieval in the context of fashion applications. We propose a unified Joint Visual Semantic Matching (JVSM) model that learns image-text compositional embeddings by jointly associating visual and textual modalities in a shared discriminative embedding space via compositional losses. JVSM has been designed with versatility and flexibility in mind, being able to perform multiple image and text tasks in a single model, such as text-image matching and language-guided retrieval. We show the effectiveness of our approach in the fashion domain, where it is difficult to express keyword-based queries given the complex specificity of fashion terms. Our experiments on three datasets (Fashion-200k, UT-Zap50k, and Fashion-iq) show that JVSM achieves state-of-the-art results on language-guided retrieval and additionally we show its capabilities to perform image and text retrieval. "

High-quality video inpainting that completes missing regions in video frames is a promising yet challenging task. State-of-the-art approaches adopt attention models to complete a frame by searching missing contents from reference frames, and further complete whole videos frame by frame. However, these approaches can suffer from inconsistent attention results along spatial and temporal dimensions, which often leads to blurriness and temporal artifacts in videos. In this paper, we propose to learn a joint Spatial-Temporal Transformer Network (STTN) for video inpainting. Specifically, we simultaneously fill missing regions in all input frames by self-attention, and propose to optimize STTN by a spatial-temporal adversarial loss. To show the superiority of the proposed model, we conduct both quantitative and qualitative evaluations by using standard stationary masks and more realistic moving object masks. Demo videos are available at https://github.com/researchmm/STTN."

Point clouds are an increasingly relevant data type but they are often corrupted by noise. We propose a deep neural network based on graph-convolutional layers that can elegantly deal with the permutation-invariance problem encountered by learning-based point cloud processing methods. The network is fully-convolutional and can build complex hierarchies of features by dynamically constructing neighborhood graphs from similarity among the high-dimensional feature representations of the points. When coupled with a loss promoting proximity to the ideal surface, the proposed approach significantly outperforms state-of-the-art methods on a variety of metrics. In particular, it is able to improve in terms of Chamfer measure and of quality of the surface normals that can be estimated from the denoised data. We also show that it is especially robust both at high noise levels and in presence of structured noise such as the one encountered in real LiDAR scans."

In this work, we propose a novel technique to generate shapes from point cloud data. A point cloud can be viewed as samples from a distribution of 3D points whose density is concentrated near the surface of the shape. Point cloud generation thus amounts to moving randomly sampled points to high-density areas. We generate point clouds by performing stochastic gradient ascent on an unnormalized probability density, thereby moving sampled points toward the high-likelihood regions. Our model directly predicts the gradient of the log density field and can be trained with a simple objective adapted from score-based generative models. We show that our method can reach state-of-the-art performance for point cloud auto-encoding and generation, while also allowing for extraction of a high-quality implicit surface."