ICCV 2019 Papers

The performance of multi-task learning in Convolutional Neural Networks (CNNs) hinges on the design of feature sharing between tasks within the architecture. The number of possible sharing patterns are combinatorial in the depth of the network and the number of tasks, and thus hand-crafting an architecture, purely based on the human intuitions of task relationships can be time-consuming and suboptimal. In this paper, we present a probabilistic approach to learning task-specific and shared representations in CNNs for multi-task learning. Specifically, we propose "stochastic filter groups" (SFG), a mechanism to assign convolution kernels in each layer to "specialist" and "generalist" groups, which are specific to and shared across different tasks, respectively. The SFG modules determine the connectivity between layers and the structures of task-specific and shared representations in the network. We employ variational inference to learn the posterior distribution over the possible grouping of kernels and network parameters. Experiments demonstrate the proposed method generalises across multiple tasks and shows improved performance over baseline methods.

As continuous-wave time-of-flight (C-ToF) cameras become popular in 3D imaging applications, they need to contend with the problem of multi-camera interference (MCI). In a multi-camera environment, a ToF camera may receive light from the sources of other cameras, resulting in large depth errors. In this paper, we propose stochastic exposure coding (SEC), a novel approach for mitigating. SEC involves dividing a camera's integration time into multiple slots, and switching the camera off and on stochastically during each slot. This approach has two benefits. First, by appropriately choosing the on probability for each slot, the camera can effectively filter out both the AC and DC components of interfering signals, thereby mitigating depth errors while also maintaining high signal-to-noise ratio. This enables high accuracy depth recovery with low power consumption. Second, this approach can be implemented without modifying the C-ToF camera's coding functions, and thus, can be used with a wide range of cameras with minimal changes. We demonstrate the performance benefits of SEC with theoretical analysis, simulations and real experiments, across a wide range of imaging scenarios.

We introduce a method for the generation of images from an input scene graph. The method separates between a layout embedding and an appearance embedding. The dual embedding leads to generated images that better match the scene graph, have higher visual quality, and support more complex scene graphs. In addition, the embedding scheme supports multiple and diverse output images per scene graph, which can be further controlled by the user. We demonstrate two modes of per-object control: (i) importing elements from other images, and (ii) navigation in the object space by selecting an appearance archetype. Our code is publicly available at https://www.github.com/ashual/scene_generation.

Rendering bridges the gap between 2D vision and 3D scenes by simulating the physical process of image formation. By inverting such renderer, one can think of a learning approach to infer 3D information from 2D images. However, standard graphics renderers involve a fundamental discretization step called rasterization, which prevents the rendering process to be differentiable, hence able to be learned. Unlike the state-of-the-art differentiable renderers, which only approximate the rendering gradient in the back propagation, we propose a truly differentiable rendering framework that is able to (1) directly render colorized mesh using differentiable functions and (2) back-propagate efficient supervision signals to mesh vertices and their attributes from various forms of image representations, including silhouette, shading and color images. The key to our framework is a novel formulation that views rendering as an aggregation function that fuses the probabilistic contributions of all mesh triangles with respect to the rendered pixels. Such formulation enables our framework to flow gradients to the occluded and far-range vertices, which cannot be achieved by the previous state-of-the-arts. We show that by using the proposed renderer, one can achieve significant improvement in 3D unsupervised single-view reconstruction both qualitatively and quantitatively. Experiments also demonstrate that our approach is able to handle the challenging tasks in image-based shape fitting, which remain nontrivial to existing differentiable renderers. Code is available at https://github.com/ShichenLiu/SoftRas.

We present SlowFast networks for video recognition. Our model involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast pathway, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition. Our models achieve strong performance for both action classification and detection in video, and large improvements are pin-pointed as contributions by our SlowFast concept. We report state-of-the-art accuracy on major video recognition benchmarks, Kinetics, Charades and AVA. Code has been made available at: https://github.com/facebookresearch/SlowFast.

We introduce SinGAN, an unconditional generative model that can be learned from a single natural image. Our model is trained to capture the internal distribution of patches within the image, and is then able to generate high quality, diverse samples that carry the same visual content as the image. SinGAN contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. This allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image. In contrast to previous single image GAN schemes, our approach is not limited to texture images, and is not conditional (i.e. it generates samples from noise). User studies confirm that the generated samples are commonly confused to be real images. We illustrate the utility of SinGAN in a wide range of image manipulation tasks.

Knowledge distillation is a widely applicable technique for training a student neural network under the guidance of a trained teacher network. For example, in neural network compression, a high-capacity teacher is distilled to train a compact student; in privileged learning, a teacher trained with privileged data is distilled to train a student without access to that data. The distillation loss determines how a teacher's knowledge is captured and transferred to the student. In this paper, we propose a new form of knowledge distillation loss that is inspired by the observation that semantically similar inputs tend to elicit similar activation patterns in a trained network. Similarity-preserving knowledge distillation guides the training of a student network such that input pairs that produce similar (dissimilar) activations in the teacher network produce similar (dissimilar) activations in the student network. In contrast to previous distillation methods, the student is not required to mimic the representation space of the teacher, but rather to preserve the pairwise similarities in its own representation space. Experiments on three public datasets demonstrate the potential of our approach.

Siamese networks are non-linear deep models that have found their ways into a broad set of problems in learning theory, thanks to their embedding capabilities. In this paper, we study Siamese networks from a new perspective and question the validity of their training procedure. We show that in the majority of cases, the objective of a Siamese network is endowed with an invariance property. Neglecting the invariance property leads to a hindrance in training the Siamese networks. To alleviate this issue, we propose two Riemannian structures and generalize a well-established accelerated stochastic gradient descent method to take into account the proposed Riemannian structures. Our empirical evaluations suggest that by making use of the Riemannian geometry, we achieve state-of-the-art results against several algorithms for the challenging problem of fine-grained image classification.

Deep learning with 3D data has progressed significantly since the introduction of convolutional neural networks that can handle point order ambiguity in point cloud data. While being able to achieve good accuracies in various scene understanding tasks, previous methods often have low training speed and complex network architecture. In this paper, we address these problems by proposing an efficient end-to-end permutation invariant convolution for point cloud deep learning. Our simple yet effective convolution operator named ShellConv uses statistics from concentric spherical shells to define representative features and resolve the point order ambiguity, allowing traditional convolution to perform on such features. Based on ShellConv we further build an efficient neural network named ShellNet to directly consume the point clouds with larger receptive fields while maintaining less layers. We demonstrate the efficacy of ShellNet by producing state-of-the-art results on object classification, object part segmentation, and semantic scene segmentation while keeping the network very fast to train.

Instance segmentation aims to detect and segment individual objects in a scene. Most existing methods rely on precise mask annotations of every category. However, it is difficult and costly to segment objects in novel categories because a large number of mask annotations is required. We introduce ShapeMask, which learns the intermediate concept of object shape to address the problem of generalization in instance segmentation to novel categories. ShapeMask starts with a bounding box detection and gradually refines it by first estimating the shape of the detected object through a collection of shape priors. Next, ShapeMask refines the coarse shape into an instance level mask by learning instance embeddings. The shape priors provide a strong cue for object-like prediction, and the instance embeddings model the instance specific appearance information. ShapeMask significantly outperforms the state-of-the-art by 6.4 and 3.8 AP when learning across categories, and obtains competitive performance in the fully supervised setting. It is also robust to inaccurate detections, decreased model capacity, and small training data. Moreover, it runs efficiently with 150ms inference time on a GPU and trains within 11 hours on TPUs. With a larger backbone model, ShapeMask increases the gap with state-of-the-art to 9.4 and 6.2 AP across categories. Code will be publicly available at: https://sites.google.com/view/shapemask/home.

Video objection detection (VID) has been a rising research direction in recent years. A central issue of VID is the appearance degradation of video frames caused by fast motion. This problem is essentially ill-posed for a single frame. Therefore, aggregating features from other frames becomes a natural choice. Existing methods rely heavily on optical flow or recurrent neural networks for feature aggregation. However, these methods emphasize more on the temporally nearby frames. In this work, we argue that aggregating features in the full-sequence level will lead to more discriminative and robust features for video object detection. To achieve this goal, we devise a novel Sequence Level Semantics Aggregation (SELSA) module. We further demonstrate the close relationship between the proposed method and the classic spectral clustering method, providing a novel view for understanding the VID problem. We test the proposed method on the ImageNet VID and the EPIC KITCHENS dataset and achieve new state-of-the-art results. Our method does not need complicated postprocessing methods such as Seq-NMS or Tubelet rescoring, which keeps the pipeline simple and clean.

Monocular face reconstruction is a challenging task in computer vision, which aims to recover 3D face geometry from a single RGB face image. Recently, deep learning based methods have achieved great improvements on monocular face reconstruction. However, for deep learning-based methods to reach optimal performance, it is paramount to have large-scale training images with ground-truth 3D face geometry, which is generally difficult for human to annotate. To tackle this problem, we propose a semi-supervised monocular reconstruction method, which jointly optimizes a shape-preserved domain-transfer CycleGAN and a shape estimation network. The framework is semi-supervised trained with 3D rendered images with ground-truth shapes and in-the-wild face images without any extra annotation. The CycleGAN network transforms all realistic images to have the rendered style and is end-to-end trained within the overall framework. This is the key difference compared with existing CycleGAN-based learning methods, which just used CycleGAN as a separate training sample generator. Novel landmark consistency loss and edge-aware shape estimation loss are proposed for our two networks to jointly solve the challenging face reconstruction problem. Extensive experiments on public face reconstruction datasets demonstrate the effectiveness of our overall method as well as the individual components.

In this work, we propose a simple yet effective semi-supervised learning approach called Augmented Distribution Alignment. We reveal that an essential sampling bias exists in semi-supervised learning due to the limited number of labeled samples, which often leads to a considerable empirical distribution mismatch between labeled data and unlabeled data. To this end, we propose to align the empirical distributions of labeled and unlabeled data to alleviate the bias. On one hand, we adopt an adversarial training strategy to minimize the distribution distance between labeled and unlabeled data as inspired by domain adaptation works. On the other hand, to deal with the small sample size issue of labeled data, we also propose a simple interpolation strategy to generate pseudo training samples. Those two strategies can be easily implemented into existing deep neural networks. We demonstrate the effectiveness of our proposed approach on the benchmark SVHN and CIFAR10 datasets. Our code is available at https://github.com/qinenergy/adanet .

Semantic scene understanding is important for various applications. In particular, self-driving cars need a fine-grained understanding of the surfaces and objects in their vicinity. Light detection and ranging (LiDAR) provides precise geometric information about the environment and is thus a part of the sensor suites of almost all self-driving cars. Despite the relevance of semantic scene understanding for this application, there is a lack of a large dataset for this task which is based on an automotive LiDAR. In this paper, we introduce a large dataset to propel research on laser-based semantic segmentation. We annotated all sequences of the KITTI Vision Odometry Benchmark and provide dense point-wise annotations for the complete 360-degree field-of-view of the employed automotive LiDAR. We propose three benchmark tasks based on this dataset: (i) semantic segmentation of point clouds using a single scan, (ii) semantic segmentation using multiple past scans, and (iii) semantic scene completion, which requires to anticipate the semantic scene in the future. We provide baseline experiments and show that there is a need for more sophisticated models to efficiently tackle these tasks. Our dataset opens the door for the development of more advanced methods, but also provides plentiful data to investigate new research directions.

Deep learning-based object detectors have shown remarkable improvements. However, supervised learning-based methods perform poorly when the train data and the test data have different distributions. To address the issue, domain adaptation transfers knowledge from the label-sufficient domain (source domain) to the label-scarce domain (target domain). Self-training is one of the powerful ways to achieve domain adaptation since it helps class-wise domain adaptation. Unfortunately, a naive approach that utilizes pseudo-labels as ground-truth degenerates the performance due to incorrect pseudo-labels. In this paper, we introduce a weak self-training (WST) method and adversarial background score regularization (BSR) for domain adaptive one-stage object detection. WST diminishes the adverse effects of inaccurate pseudo-labels to stabilize the learning procedure. BSR helps the network extract discriminative features for target backgrounds to reduce the domain shift. Two components are complementary to each other as BSR enhances discrimination between foregrounds and backgrounds, whereas WST strengthen class-wise discrimination. Experimental results show that our approach effectively improves the performance of the one-stage object detection in unsupervised domain adaptation setting.

Domain adaptation in person re-identification (re-ID) has always been a challenging task. In this work, we explore how to harness the similar natural characteristics existing in the samples from the target domain for learning to conduct person re-ID in an unsupervised manner. Concretely, we propose a Self-similarity Grouping (SSG) approach, which exploits the potential similarity (from the global body to local parts) of unlabeled samples to build multiple clusters from different views automatically. These independent clusters are then assigned with labels, which serve as the pseudo identities to supervise the training process. We repeatedly and alternatively conduct such a grouping and training process until the model is stable. Despite the apparent simplify, our SSG outperforms the state-of-the-arts by more than 4.6% (DukeMTMC-Market1501) and 4.4% (Market1501-DukeMTMC) in mAP, respectively. Upon our SSG, we further introduce a clustering-guided semisupervised approach named SSG ++ to conduct the one-shot domain adaption in an open set setting (i.e. the number of independent identities from the target domain is unknown). Without spending much effort on labeling, our SSG ++ can further promote the mAP upon SSG by 10.7% and 6.9%, respectively. Our Code is available at: https://github.com/OasisYang/SSG .

Despite the success of Generative Adversarial Networks (GANs), mode collapse remains a serious issue during GAN training. To date, little work has focused on understanding and quantifying which modes have been dropped by a model. In this work, we visualize mode collapse at both the distribution level and the instance level. First, we deploy a semantic segmentation network to compare the distribution of segmented objects in the generated images with the target distribution in the training set. Differences in statistics reveal object classes that are omitted by a GAN. Second, given the identified omitted object classes, we visualize the GAN's omissions directly. In particular, we compare specific differences between individual photos and their approximate inversions by a GAN. To this end, we relax the problem of inversion and solve the tractable problem of inverting a GAN layer instead of the entire generator. Finally, we use this framework to analyze several recent GANs trained on multiple datasets and identify their typical failure cases.

Deep learning has recently been applied with impressive results to extreme low-light imaging. Despite the success of single-image processing, extreme low-light video processing is still intractable due to the difficulty of collecting raw video data with corresponding ground truth. Collecting long-exposure ground truth, as was done for single-image processing, is not feasible for dynamic scenes. In this paper, we present deep processing of very dark raw videos: on the order of one lux of illuminance. To support this line of work, we collect a new dataset of raw low-light videos, in which high-resolution raw data is captured at video rate. At this level of darkness, the signal-to-noise ratio is extremely low (negative if measured in dB) and the traditional image processing pipeline generally breaks down. A new method is presented to address this challenging problem. By carefully designing a learning-based pipeline and introducing a new loss function to encourage temporal stability, we train a siamese network on static raw videos, for which ground truth is available, such that the network generalizes to videos of dynamic scenes at test time. Experimental results demonstrate that the presented approach outperforms state-of-the-art models for burst processing, per-frame processing, and blind temporal consistency.

We present the next generation of MobileNets based on a combination of complementary search techniques as well as a novel architecture design. MobileNetV3 is tuned to mobile phone CPUs through a combination of hardware-aware network architecture search (NAS) complemented by the NetAdapt algorithm and then subsequently improved through novel architecture advances. This paper starts the exploration of how automated search algorithms and network design can work together to harness complementary approaches improving the overall state of the art. Through this process we create two new MobileNet models for release: MobileNetV3-Large and MobileNetV3-Small which are targeted for high and low resource use cases. These models are then adapted and applied to the tasks of object detection and semantic segmentation. For the task of semantic segmentation (or any dense pixel prediction), we propose a new efficient segmentation decoder Lite Reduced Atrous Spatial Pyramid Pooling (LR-ASPP). We achieve new state of the art results for mobile classification, detection and segmentation. MobileNetV3-Large is 3.2% more accurate on ImageNet classification while reducing latency by 20% compared to MobileNetV2. MobileNetV3-Small is 6.6% more accurate compared to a MobileNetV2 model with comparable latency. MobileNetV3-Large detection is over 25% faster at roughly the same accuracy as MobileNetV2 on COCO detection. MobileNetV3-Large LR-ASPP is 34% faster than MobileNetV2 R-ASPP at similar accuracy for Cityscapes segmentation.