ICCV 2019 Papers

Learning an effective similarity measure between image representations is key to the success of recent advances in visual search tasks (e.g. verification or zero-shot learning). Although the metric learning part is well addressed, this metric is usually computed over the average of the extracted deep features. This representation is then trained to be discriminative. However, these deep features tend to be scattered across the feature space. Consequently, the representations are not robust to outliers, object occlusions, background variations, etc. In this paper, we tackle this scattering problem with a distribution-aware regularization named HORDE. This regularizer enforces visually-close images to have deep features with the same distribution which are well localized in the feature space. We provide a theoretical analysis supporting this regularization effect. We also show the effectiveness of our approach by obtaining state-of-the-art results on 4 well-known datasets (Cub-200-2011, Cars-196, Stanford Online Products and Inshop Clothes Retrieval).

Understanding dynamic 3D environment is crucial for robotic agents and many other applications. We propose a novel neural network architecture called MeteorNet for learning representations for dynamic 3D point cloud sequences. Different from previous work that adopts a grid-based representation and applies 3D or 4D convolutions, our network directly processes point clouds. We propose two ways to construct spatiotemporal neighborhoods for each point in the point cloud sequence. Information from these neighborhoods is aggregated to learn features per point. We benchmark our network on a variety of 3D recognition tasks including action recognition, semantic segmentation and scene flow estimation. MeteorNet shows stronger performance than previous grid-based methods while achieving state-of-the-art performance on Synthia. MeteorNet also outperforms previous baseline methods that are able to process at most two consecutive point clouds. To the best of our knowledge, this is the first work on deep learning for dynamic raw point cloud sequences.

In this paper, we propose a novel meta learning approach for automatic channel pruning of very deep neural networks. We first train a PruningNet, a kind of meta network, which is able to generate weight parameters for any pruned structure given the target network. We use a simple stochastic structure sampling method for training the PruningNet. Then, we apply an evolutionary procedure to search for good-performing pruned networks. The search is highly efficient because the weights are directly generated by the trained PruningNet and we do not need any finetuning at search time. With a single PruningNet trained for the target network, we can search for various Pruned Networks under different constraints with little human participation. Compared to the state-of-the-art pruning methods, we have demonstrated superior performances on MobileNet V1/V2 and ResNet. Codes are available on https://github.com/liuzechun/MetaPruning.

Training models to high-end performance requires availability of large labeled datasets, which are expensive to get. The goal of our work is to automatically synthesize labeled datasets that are relevant for a downstream task. We propose Meta-Sim, which learns a generative model of synthetic scenes, and obtain images as well as its corresponding ground-truth via a graphics engine. We parametrize our dataset generator with a neural network, which learns to modify attributes of scene graphs obtained from probabilistic scene grammars, so as to minimize the distribution gap between its rendered outputs and target data. If the real dataset comes with a small labeled validation set, we additionally aim to optimize a meta-objective, i.e. downstream task performance. Experiments show that the proposed method can greatly improve content generation quality over a human-engineered probabilistic scene grammar, both qualitatively and quantitatively as measured by performance on a downstream task.

Few-shot learning, i.e., learning novel concepts from few examples, is fundamental to practical visual recognition systems. While most of existing work has focused on few-shot classification, we make a step towards few-shot object detection, a more challenging yet under-explored task. We develop a conceptually simple but powerful meta-learning based framework that simultaneously tackles few-shot classification and few-shot localization in a unified, coherent way. This framework leverages meta-level knowledge about "model parameter generation" from base classes with abundant data to facilitate the generation of a detector for novel classes. Our key insight is to disentangle the learning of category-agnostic and category-specific components in a CNN based detection model. In particular, we introduce a weight prediction meta-model that enables predicting the parameters of category-specific components from few examples. We systematically benchmark the performance of modern detectors in the small-sample size regime. Experiments in a variety of realistic scenarios, including within-domain, cross-domain, and long-tailed settings, demonstrate the effectiveness and generality of our approach under different notions of novel classes.

Resembling the rapid learning capability of human, low-shot learning empowers vision systems to understand new concepts by training with few samples. Leading approaches derived from meta-learning on images with a single visual object. Obfuscated by a complex background and multiple objects in one image, they are hard to promote the research of low-shot object detection/segmentation. In this work, we present a flexible and general methodology to achieve these tasks. Our work extends Faster /Mask R-CNN by proposing meta-learning over RoI (Region-of-Interest) features instead of a full image feature. This simple spirit disentangles multi-object information merged with the background, without bells and whistles, enabling Faster /Mask R-CNN turn into a meta-learner to achieve the tasks. Specifically, we introduce a Predictor-head Remodeling Network (PRN) that shares its main backbone with Faster /Mask R-CNN. PRN receives images containing low-shot objects with their bounding boxes or masks to infer their class attentive vectors. The vectors take channel-wise soft-attention on RoI features, remodeling those R-CNN predictor heads to detect or segment the objects consistent with the classes these vectors represent. In our experiments, Meta R-CNN yields the new state of the art in low-shot object detection and improves low-shot object segmentation by Mask R-CNN. Code: https://yanxp.github.io/metarcnn.html.

Rapid advances in 2D perception have led to systems that accurately detect objects in real-world images. However, these systems make predictions in 2D, ignoring the 3D structure of the world. Concurrently, advances in 3D shape prediction have mostly focused on synthetic benchmarks and isolated objects. We unify advances in these two areas. We propose a system that detects objects in real-world images and produces a triangle mesh giving the full 3D shape of each detected object. Our system, called Mesh R-CNN, augments Mask R-CNN with a mesh prediction branch that outputs meshes with varying topological structure by first predicting coarse voxel representations which are converted to meshes and refined with a graph convolution network operating over the mesh's vertices and edges. We validate our mesh prediction branch on ShapeNet, where we outperform prior work on single-image shape prediction. We then deploy our full Mesh R-CNN system on Pix3D, where we jointly detect objects and predict their 3D shapes.

Learning discriminative image feature embeddings is of great importance to visual recognition. To achieve better feature embeddings, most current methods focus on designing different network structures or loss functions, and the estimated feature embeddings are usually only related to the input images. In this paper, we propose Memory-based Neighbourhood Embedding (MNE) to enhance a general CNN feature by considering its neighbourhood. The method aims to solve two critical problems, i.e., how to acquire more relevant neighbours in the network training and how to aggregate the neighbourhood information for a more discriminative embedding. We first augment an episodic memory module into the network, which can provide more relevant neighbours for both training and testing. Then the neighbours are organized in a tree graph with the target instance as the root node. The neighbourhood information is gradually aggregated to the root node in a bottom-up manner, and aggregation weights are supervised by the class relationships between the nodes. We apply MNE on image search and few shot learning tasks. Extensive ablation studies demonstrate the effectiveness of each component, and our method significantly outperforms the state-of-the-art approaches.

Deep autoencoder has been extensively used for anomaly detection. Training on the normal data, the autoencoder is expected to produce higher reconstruction error for the abnormal inputs than the normal ones, which is adopted as a criterion for identifying anomalies. However, this assumption does not always hold in practice. It has been observed that sometimes the autoencoder "generalizes" so well that it can also reconstruct anomalies well, leading to the miss detection of anomalies. To mitigate this drawback for autoencoder based anomaly detector, we propose to augment the autoencoder with a memory module and develop an improved autoencoder called memory-augmented autoencoder, i.e. MemAE. Given an input, MemAE firstly obtains the encoding from the encoder and then uses it as a query to retrieve the most relevant memory items for reconstruction. At the training stage, the memory contents are updated and are encouraged to represent the prototypical elements of the normal data. At the test stage, the learned memory will be fixed, and the reconstruction is obtained from a few selected memory records of the normal data. The reconstruction will thus tend to be close to a normal sample. Thus the reconstructed errors on anomalies will be strengthened for anomaly detection. MemAE is free of assumptions on the data type and thus general to be applied to different tasks. Experiments on various datasets prove the excellent generalization and high effectiveness of the proposed MemAE.

Deep hashing enables computation and memory efficient image search through end-to-end learning of feature representations and binary codes. While linear scan over binary hash codes is more efficient than over the high-dimensional representations, its linear-time complexity is still unacceptable for very large databases. Hamming space retrieval enables constant-time search through hash lookups, where for each query, there is a Hamming ball centered at the query and the data points within the ball are returned as relevant. Since inside the Hamming ball implies retrievable while outside irretrievable, it is crucial to explicitly characterize the Hamming ball. The main idea of this work is to directly embody the Hamming radius into the loss functions, leading to Maximum-Margin Hamming Hashing (MMHH), a new model specifically optimized for Hamming space retrieval. We introduce a max-margin t-distribution loss, where the t-distribution concentrates more similar data points to be within the Hamming ball, and the margin characterizes the Hamming radius such that less penalization is applied to similar data points within the Hamming ball. The loss function also introduces robustness to data noise, where the similarity supervision may be inaccurate in practical problems. The model is trained end-to-end using a new semi-batch optimization algorithm tailored to extremely imbalanced data. Our method yields state-of-the-art results on four datasets and shows superior performance on noisy data.

This paper presents a new method for shadow removal using unpaired data, enabling us to avoid tedious annotations and obtain more diverse training samples. However, directly employing adversarial learning and cycle-consistency constraints is insufficient to learn the underlying relationship between the shadow and shadow-free domains, since the mapping between shadow and shadow-free images is not simply one-to-one. To address the problem, we formulate Mask-ShadowGAN, a new deep framework that automatically learns to produce a shadow mask from the input shadow image and then takes the mask to guide the shadow generation via re-formulated cycle-consistency constraints. Particularly, the framework simultaneously learns to produce shadow masks and learns to remove shadows, to maximize the overall performance. Also, we prepared an unpaired dataset for shadow removal and demonstrated the effectiveness of Mask-ShadowGAN on various experiments, even it was trained on unpaired data.

Pedestrian detection relying on deep convolution neural networks has made significant progress. Though promising results have been achieved on standard pedestrians, the performance on heavily occluded pedestrians remains far from satisfactory. The main culprits are intra-class occlusions involving other pedestrians and inter-class occlusions caused by other objects, such as cars and bicycles. These results in a multitude of occlusion patterns. We propose an approach for occluded pedestrian detection with the following contributions. First, we introduce a novel mask-guided attention network that fits naturally into popular pedestrian detection pipelines. Our attention network emphasizes on visible pedestrian regions while suppressing the occluded ones by modulating full body features. Second, we empirically demonstrate that coarse-level segmentation annotations provide reasonable approximation to their dense pixel-wise counterparts. Experiments are performed on CityPersons and Caltech datasets. Our approach sets a new state-of-the-art on both datasets. Our approach obtains an absolute gain of 9.5% in log-average miss rate, compared to the best reported results [32] on the heavily occluded HO pedestrian set of CityPersons test set. Further, on the HO pedestrian set of Caltech dataset, our method achieves an absolute gain of 5.0% in log-average miss rate, compared to the best reported results [13]. Code and models are available at: https://github.com/Leotju/MGAN.

Capturing human motion in natural scenarios means moving motion capture out of the lab and into the wild. Typical approaches rely on fixed, calibrated, cameras and reflective markers on the body, significantly limiting the motions that can be captured. To make motion capture truly unconstrained, we describe the first fully autonomous outdoor capture system based on flying vehicles. We use multiple micro-aerial-vehicles(MAVs), each equipped with a monocular RGB camera, an IMU, and a GPS receiver module. These detect the person, optimize their position, and localize themselves approximately. We then develop a markerless motion capture method that is suitable for this challenging scenario with a distant subject, viewed from above, with approximately calibrated and moving cameras. We combine multiple state-of-the-art 2D joint detectors with a 3D human body model and a powerful prior on human pose. We jointly optimize for 3D body pose and camera pose to robustly fit the 2D measurements. To our knowledge, this is the first successful demonstration of outdoor, full-body, markerless motion capture from autonomous flying vehicles.

Typical multi-task learning (MTL) methods rely on architectural adjustments and a large trainable parameter set to jointly optimize over several tasks. However, when the number of tasks increases so do the complexity of the architectural adjustments and resource requirements. In this paper, we introduce a method which applies a conditional feature-wise transformation over the convolutional activations that enables a model to successfully perform a large number of tasks. To distinguish from regular MTL, we introduce Many Task Learning (MaTL) as a special case of MTL where more than 20 tasks are performed by a single model. Our method dubbed Task Routing (TR) is encapsulated in a layer we call the Task Routing Layer (TRL), which applied in an MaTL scenario successfully fits hundreds of classification tasks in one model. We evaluate on 5 datasets and the Visual Decathlon (VD) challenge against strong baselines and state-of-the-art approaches.

Understanding people's actions and interactions typically depends on seeing them. Automating the process of action recognition from visual data has been the topic of much research in the computer vision community. But what if it is too dark, or if the person is occluded or behind a wall? In this paper, we introduce a neural network model that can detect human actions through walls and occlusions, and in poor lighting conditions. Our model takes radio frequency (RF) signals as input, generates 3D human skeletons as an intermediate representation, and recognizes actions and interactions of multiple people over time. By translating the input to an intermediate skeleton-based representation, our model can learn from both vision-based and RF-based datasets, and allow the two tasks to help each other. We show that our model achieves comparable accuracy to vision-based action recognition systems in visible scenarios, yet continues to work accurately when people are not visible, hence addressing scenarios that are beyond the limit of today's vision-based action recognition.

We study the multi-round response generation in visual dialog, where a response is generated according to a visually grounded conversational history. Given a triplet: an image, Q&A history, and current question, all the prevailing methods follow a codec (i.e., encoder-decoder) fashion in a supervised learning paradigm: a multimodal encoder encodes the triplet into a feature vector, which is then fed into the decoder for the current answer generation, supervised by the ground-truth. However, this conventional supervised learning does NOT take into account the impact of imperfect history, violating the conversational nature of visual dialog and thus making the codec more inclined to learn history bias but not contextual reasoning. To this end, inspired by the actor-critic policy gradient in reinforcement learning, we propose a novel training paradigm called History Advantage Sequence Training (HAST). Specifically, we intentionally impose wrong answers in the history, obtaining an adverse critic, and see how the historic error impacts the codec's future behavior by History Advantage -- a quantity obtained by subtracting the adverse critic from the gold reward of ground-truth history. Moreover, to make the codec more sensitive to the history, we propose a novel attention network called History-Aware Co-Attention Network (HACAN) which can be effectively trained by using HAST. Experimental results on three benchmarks: VisDial v0.9&v1.0 and GuessWhat?!, show that the proposed HAST strategy consistently outperforms the state-of-the-art supervised counterparts.

Recent studies have shown remarkable success in face manipulation task with the advance of GANs and VAEs paradigms, but the outputs are sometimes limited to low-resolution and lack of diversity. In this work, we propose Additive Focal Variational Auto-encoder (AF-VAE), a novel approach that can arbitrarily manipulate high-resolution face images using a simple yet effective model and only weak supervision of reconstruction and KL divergence losses. First, a novel additive Gaussian Mixture assumption is introduced with an unsupervised clustering mechanism in the structural latent space, which endows better disentanglement and boosts multi-modal representation with external memory. Second, to improve the perceptual quality of synthesized results, two simple strategies in architecture design are further tailored and discussed on the behavior of Human Visual System (HVS) for the first time, allowing for fine control over the model complexity and sample quality. Human opinion studies and new state-of-the-art Inception Score (IS) / Frechet Inception Distance (FID) demonstrate the superiority of our approach over existing algorithms, advancing both the fidelity and extremity of face manipulation task.

We present a deep-learning architecture for multi-view stereo with conditional random fields (MVSCRF). Given an arbitrary number of input images, we first use a U-shape neural network to extract deep features incorporating both global and local information, and then build a 3D cost volume for the reference camera. Unlike previous learning based methods, we explicitly constraint the smoothness of depth maps by using conditional random fields (CRFs) after the stage of cost volume regularization. The CRFs module is implemented as recurrent neural networks so that the whole pipeline can be trained end-to-end. Our results show that the proposed pipeline outperforms previous state-of-the-arts on large-scale DTU dataset. We also achieve comparable results with state-of-the-art learning based methods on outdoor Tanks and Temples dataset without fine-tuning, which demonstrates our method's generalization ability.

How to correctly stress hard samples in metric learning is critical for visual recognition tasks, especially in challenging person re-ID applications. Pedestrians across cameras with significant appearance variations are easily confused, which could bias the learned metric and slow down the convergence rate. In this paper, we propose a novel weighted complete bipartite graph based maximum-value perfect (MVP) matching for mining the hard samples from a batch of samples. It can emphasize the hard positive and negative sample pairs respectively, and thus relieve adverse optimization and sample imbalance problems. We then develop a new batch-wise MVP matching based loss objective and combine it in an end-to-end deep metric learning manner. It leads to significant improvements in both convergence rate and recognition performance. Extensive empirical results on five person re-ID benchmark datasets, i.e., Market-1501, CUHK03-Detected, CUHK03-Labeled, Duke-MTMC, and MSMT17, demonstrate the superiority of the proposed method. It can accelerate the convergence rate significantly while achieving state-of-the-art performance. The source code of our method is available at https://github.com/IAAI-CVResearchGroup/MVP-metric.