WACV 2021 Papers

We present a system for real-time RGBD-based estimation of 3D human pose. We use parametric 3D deformable human mesh model (SMPL-X) as a representation and focus on the real-time estimation of parameters for the body pose, hands pose and facial expression from Kinect Azure RGB-D camera. We train estimators of body pose and facial expression parameters. Both estimators use previously published landmark extractors as input and custom annotated datasets for supervision, while hand pose is estimated directly by a previously published method. We combine the predictions of those estimators into a temporally-smooth human pose. We train the facial expression extractor on a large talking face dataset, which we annotate with facial expression parameters. For the body pose we collect and annotate a dataset of 56 people captured from a rig of 5 Kinect Azure RGB-D cameras and use it together with a large motion capture AMASS dataset. Our RGB-D body pose model outperforms the state-of-the-art RGB-only methods and works on the same level of accuracy compared to a slower RGB-D optimization-based solution. The combined system runs at 25 FPS on a server with a single GPU. The code will be available at saic-violet.github.io/rgbd-kinect-pose

We present a novel algorithm for transferring artistic styles of semantically meaningful local regions of an image onto local regions of a target video while preserving its photorealism. Local regions may be selected either fully automatically from an image, through using video segmentation algorithms, or from casual user guidance such as scribbles. Our method, based on a deep neural network architecture inspired by recent work in photorealistic style transfer, is real-time and works on arbitrary inputs without runtime optimization once trained on a diverse dataset of artistic styles. By augmenting our video dataset with noisy semantic labels and jointly optimizing over style, content, mask, and temporal losses, our method can cope with a variety of imperfections in the input and produce temporally coherent videos without visual artifacts. We demonstrate our method on a variety of style images and target videos, including the ability to transfer different styles onto multiple objects simultaneously, and smoothly transition between styles in time.

In this paper, we propose a unified real-time framework for gait-based age estimation and gender classification that uses just a single image, which reduces the latency in video capturing compared with the existing methods based on a gait cycle. To cope with the problem of lacking motion information in the input single image, we first reconstruct a gait cycle of a silhouette sequence from the input image via a gait cycle reconstruction network. The reconstructed gait cycle is then fed into a state-of-the-art gait recognition network for feature representation learning, which is further used to obtain the class of the gender and the estimated probability distribution of integer age labels. Unlike the existing methods focusing on the gait sequences captured from the side view, the proposed method is applicable to the gait images from an arbitrary view with a single trained model, which is more suitable for real-world application scenarios (e.g., automatic access control). Stand-alone and client-server online systems were implemented based on the proposed method, which validates the real-time/online property in actual scenes. The experiments on the world's largest multi-view gait dataset demonstrate the effectiveness of the proposed method, which achieves performance improvement compared with the benchmark algorithms.

RarePlanes is a unique open-source machine learning dataset that incorporates both real and synthetically generated satellite imagery. The RarePlanes dataset specifically focuses on the value of synthetic data to aid computer vision algorithms in their ability to automatically detect aircraft and their attributes in satellite imagery. Although other synthetic/real combination datasets exist, RarePlanes is the largest openly-available very-high resolution dataset built to test the value of synthetic data from an overhead perspective. Previous research has shown that synthetic data can reduce the amount of real training data needed and potentially improve performance for many tasks in the computer vision domain. The real portion of the dataset consists of 253 Maxar WorldView-3 satellite scenes spanning 112 locations and 2,142 km^2 with 14,700 hand-annotated aircraft. The accompanying synthetic dataset is generated via AI.Reverie's simulation platform and features 50,000 synthetic satellite images simulating a total area of 9331.2 km^2 with 630,000 aircraft annotations. Both the real and synthetically generated aircraft feature 10 fine grain attributes including: aircraft length, wingspan, wing-shape, wing-position, wingspan class, propulsion, number of engines, number of vertical-stabilizers, presence of canards, and aircraft role. Finally, we conduct extensive experiments to evaluate the real and synthetic datasets and compare performances. By doing so, we show the value of synthetic data for the task of detecting and classifying aircraft from an overhead perspective.

Radar is usually more robust than the camera in severe driving scenarios, e.g., weak/strong lighting and bad weather. However, unlike RGB images captured by a camera, the semantic information from the radar signals is noticeably difficult to extract. In this paper, we propose a deep radar object detection network (RODNet), to effectively detect objects purely from the carefully processed radar frequency data in the format of range-azimuth frequency heatmaps (RAMaps). Three different 3D autoencoder based architectures are introduced to predict object confidence distribution from each snippet of the input RAMaps. The final detection results are then calculated using our post-processing method, called location-based non-maximum suppression (L-NMS). Instead of using burdensome human-labeled ground truth, we train the RODNet using the annotations generated automatically by a novel 3D localization method using a camera-radar fusion (CRF) strategy. To train and evaluate our method, we build a new dataset -- CRUW, containing synchronized videos and RAMaps in various driving scenarios. After intensive experiments, our RODNet shows favorable object detection performance without the presence of the camera.

Learning from a few examples is an important practical aspect of training classifiers. Various works have examined this aspect quite well. However, all existing approaches assume that the few examples provided are always correctly labeled. This is a strong assumption, especially if one considers the current techniques for labeling using crowd-based labeling services. We address this issue by proposing a novel robust few-shot learning approach. Our method relies on generating robust prototypes from a set of few examples. Specifically, our method refines the class prototypes by producing hybrid features from the support examples of each class. The refined prototypes help to classify the query images better. Our method can replace the evaluation phase of any few-shot learning method that uses a nearest neighbor prototype-based evaluation procedure to make them robust. We evaluate our method on standard mini-ImageNet and tiered-ImageNet datasets. We perform experiments with various label corruption rates in the support examples of the few-shot classes. We obtain significant improvement over widely used few-shot learning methods that suffer significant performance degeneration in the presence of label noise. We finally provide extensive ablation experiments to validate our method.

We propose a real-time general purpose semantic segmentation architecture, RGPNet, which achieves significant performance gain in complex environments. RGPNet consists of a light-weight asymmetric encoder-decoder and an adaptor. The adaptor helps preserve and refine the abstract concepts from multiple levels of distributed representations between encoder and decoder. It also facilitates the gradient flow from deeper layers to shallower layers. Our experiments demonstrate that RGPNet can generate segmentation results in real-time with comparable accuracy to the state-of-the-art non-real-time heavy models. Moreover, towards green AI, we show that using an optimized label-relaxation technique with progressive resizing can reduce the training time by up to 60% while preserving the performance. We conclude that RGPNet obtains a better speed-accuracy trade-off across multiple datasets.

Facial image inpainting is a problem that is widely studied, and in recent years the introduction of Generative Adversarial Networks, has led to improvements in the field. Unfortunately some issues persists, in particular when blending the missing pixels with the visible ones. We address the problem by proposing a Wasserstein GAN combined with a new reverse mask operator, namely Reverse Masking Network (R-MNet), a perceptual adversarial network for image inpainting. The reverse mask operator transfers the reverse masked image to the end of the encoder-decoder network leaving only valid pixels to be inpainted. Additionally, we propose a new loss function computed in feature space to target only valid pixels combined with adversarial training. These then capture data distributions and generate images similar to those in the training data with achieved realism (realistic and coherent) on the output images. We evaluate our method on publicly available dataset, and compare with state-of-the-art methods. We show that our method is able to generalize to high-resolution inpainting task, and further show more realistic outputs that are plausible to the human visual system when compared with the state-of-the-art methods. https://github.com/Jireh-Jam/R-MNET-A-Perceptual-Adversarial-Network-for-Image-Inpainting

Vision for autonomous driving is a uniquely challenging problem: the number of tasks required for full scene understanding is large and diverse; the quality requirements on each task are stringent due to the safety-critical nature of the application; and the latency budget is limited, requiring real-time solutions. In this work we address these challenges with QuadroNet, a one-shot network that jointly produces four outputs: 2D detections, instance segmentation, semantic segmentation, and monocular depth estimates in real-time (>60fps) on consumer-grade GPU hardware. On a challenging real-world autonomous driving dataset, we demonstrate an increase of +2.4% mAP for detection, +3.15% mIoU for semantic segmentation, +5.05% mAP@0.5 for instance segmentation and +1.36% in delta<1.25 for depth prediction over a baseline approach. We also compare our work against other multi-task learning approaches on Cityscapes and demonstrate state-of-the-art results.

In this paper, we focus on semi-supervised object detection to boost performance of proposal-based object detectors (a.k.a. two-stage object detectors) by training on both labeled and unlabeled data. However, it is non-trivial to train object detectors on unlabeled data due to the unavailability of ground truth labels. To address this problem, we present a proposal learning approach to learn proposal features and predictions from both labeled and unlabeled data. The approach consists of a self-supervised proposal learning module and a consistency-based proposal learning module. In the self-supervised proposal learning module, we present a proposal location loss and a contrastive loss to learn context-aware and noise-robust proposal features respectively. In the consistency-based proposal learning module, we apply consistency losses to both bounding box classification and regression predictions of proposals to learn noise-robust proposal features and predictions. Our approach enjoys the following benefits: 1) encouraging more context information to be delivered in the proposals learning procedure; 2) noisy proposal features and enforcing consistency to allow noise-robust object detection; 3) building a general and high-performance semi-supervised object detection framework, which can be easily adapted to proposal-based object detectors with different backbone architectures. Experiments are conducted on the COCO dataset with all available labeled and unlabeled data. Results demonstrate that our approach consistently improves the performance of fully-supervised baselines. In particular, after combining with data distillation [38], our approach improves AP by about 2.0% and 0.9% on average compared to fully-supervised baselines and data distillation baselines respectively.

Neural networks are highly effective tools for pose estimation. However, as in other computer vision tasks, robustness to out-of-domain data remains a challenge, especially for small training sets that are common for real-world applications. Here, we probe the generalization ability with three architecture classes (MobileNetV2s, ResNets, and EfficientNets) for pose estimation. We developed a dataset of 30 horses that allowed for both "within-domain" and "out-of-domain" (unseen horse) benchmarking---this is a crucial test for robustness that current human pose estimation benchmarks do not directly address. We show that better ImageNet-performing architectures perform better on both within- and out-of-domain data if they are first pretrained on ImageNet. We additionally show that better ImageNet models generalize better across animal species. Furthermore, we introduce Horse-C, a new benchmark for common corruptions for pose estimation, and confirm that pretraining increases performance in this domain shift context as well. Overall, our results demonstrate that transfer learning is beneficial for out-of-domain robustness.

Stochastic gradient descent (SGD) is successfully applied to train deep convolutional neural networks (CNNs) on various computer vision tasks. Since fixed step-size SGD converges to so-called error plateau, it is applied in combination with decaying learning rate to reach a favorable optimum. In this paper, we propose a simple yet effective optimization method to improve SGD with a phase-wise decay of learning rate. Through analyzing both a loss surface around the error plateau and a structure of the SGD optimization process, the proposed method is formulated to improve convergence as well as initialization at each training phase by efficiently aggregating the CNN parameters along the optimization sequence. The method keeps the simplicity of SGD while touching the SGD procedure only a few times during training. The experimental results on image classification tasks thoroughly validate the effectiveness of the proposed method in comparison to the other methods.

Despite significant progress, controlled generation of complex images with interacting people remains difficult. Existing layout generation methods fall short of synthesizing realistic person instances; while pose-guided generation approaches focus on a single person and assume simple or known backgrounds. To tackle these limitations, we propose a new problem, person in context synthesis, which aims to synthesize diverse person instance(s) in consistent contexts, with user control over both. The context is specified by the bounding box object layout which lacks shape information, while pose of the person(s) by keypoints which are sparsely annotated. To handle the stark difference in input structures, we proposed two separate neural branches to attentively composite the respective (context/person) inputs into shared "compositional structural space", which encodes shape, location and appearance information for both context and person structures in a disentangled manner. This structural space is then decoded to the image space using multi-level feature modulation strategy, and learned in a self supervised manner from image collections and their corresponding inputs. Extensive experiments on two large-scale datasets (COCO-Stuff and Visual Genome ) demonstrate that our framework outperforms state-of-the-art methods w.r.t. synthesis quality.

While object part segmentation is useful for many applications, typical approaches require a large amount of labeled data to train a model for good performance. To reduce the labeling effort, weak supervision cues such as object keypoints have been used to generate pseudo-part annotations which can subsequently be used to train larger models. However, previous weakly-supervised part segmentation methods require the same object classes during both training and testing. We propose a new model to use key-point guidance for segmenting parts of novel object classes given that they have similar structures as seen objects --different types of four-legged animals, for example. We show that a non-parametric template matching approach is more effective than pixel classification for part segmentation, especially for small or less frequent parts. To evaluate the generalizability of our approach, we introduce two new datasets that contain 200 quadrupeds in total with both key-point and part segmentation annotations. We show that our approach can outperform existing models by a large mar-gin on the novel object part segmentation task using limited part segmentation labels during training.

Image outpainting is a very intriguing problem as the outside of a given image can be continuously filled by considering as the context of the image. This task has two main challenges. The first is to maintain the spatial consistency in contents of generated regions and the original input. The second is to generate a high-quality large image with a small amount of adjacent information. Conventional image outpainting methods generate inconsistent, blurry, and repeated pixels. To alleviate the difficulty of an outpainting problem, we propose a novel image outpainting method using bidirectional boundary region rearrangement. We rearrange the image to benefit from the image inpainting task by reflecting more directional information. The bidirectional boundary region rearrangement enables the generation of the missing region using bidirectional information similar to that of the image inpainting task, thereby generating the higher quality than the conventional methods using unidirectional information. Moreover, we use the edge map generator that considers images as original input with structural information and hallucinates the edges of unknown regions to generate the image. Our proposed method is compared with other state-of-the-art outpainting and inpainting methods both qualitatively and quantitatively. We further compared and evaluated them using BRISQUE, one of the No-Reference image quality assessment (IQA) metrics, to evaluate the naturalness of the output. The experimental results demonstrate that our method outperforms other methods and generates new images with 360degpanoramic characteristics.

The human visual system can detect objects of unseen categories from merely a few examples. However, such capability remains absent in state-of-the-art detectors. To bridge this gap, several attempts have been proposed to perform few-shot detection by incorporating meta-learning techniques. Such methods can improve detection performance on unseen categories, but also add huge computational burden, and usually degrade detection performance on seen categories. In this paper, we present PNPDet, a novel Plug-and-Play Detector, for efficient few-shot detection without forgetting. It introduces a simple but effective architecture with separate sub-networks that disentangles the recognition of base and novel categories and prevents hurting performance on known categories while learning new concepts. Distance metric learning is further incorporated into sub-networks, consistently boosting detection performance for both base and novel categories. Experiments show that the proposed PNPDet can achieve comparable few-shot detection performance on unseen categories while not losing accuracy on seen categories, and also remain efficient and flexible at the same time.

Recent literature addressed the monocular 3D pose estimation task very satisfactorily. In these studies, different persons are usually treated as independent pose instances to estimate. However, in many every-day situations, people are interacting, and the pose of an individual depends on the pose of his/her interactees. In this paper, we investigate how to exploit these dependency to enhance current, and possibly future, deep networks for 3D monocular pose estimation. Our pose interacting network, or PI-Net, inputs the initial pose estimates of a variable number of interactees into a recurrent network used to refine the pose of the person-of-interest. Evaluating such a method is challenging due to the limited availability of public annotated multi-person 3D human pose datasets. We demonstrate the effectiveness of our method in the MuPoTS dataset, setting the new state-of-the-art on it. Qualitative results on other multi-person datasets (for which 3D pose ground-truth is not available) showcase the proposed PI-Net.

Handling long and complex temporal information is an important factor for action detection tasks. This challenge is further aggravated by densely distributed actions in untrimmed videos. Previous action detection methods are failing in selecting the key temporal information in videos of long length. To this end, we introduce the Dilated Attention Layer (DAL). Compared to previous temporal convolution layer, DAL allocates attentional weights to each feature in the kernel, which enables DAL to learn better local representation across time. Furthermore, DAL when accompanied by dilated kernels is able to learn a global representation of several minutes long videos which is crucial for the task of action detection. Finally, we introduce Pyramid Dilated Attention Network (PDAN) which is build upon DAL. With the help of DAL combining with dilation and residual links, PDAN can model short-term and long-term temporal relations simultaneously by focusing on local segments at the level of low and high temporal receptive fields. This property enables PDAN to handle complex temporal relations between different action instances in long untrimmed videos. To corroborate the effectiveness and robustness of our proposed method, we evaluate it on three densely annotated, multi-label datasets: MultiTHUMOS, Charades and an Inhouse dataset, outperforming the state-of-the-art results.

Deep Subspace Clustering Networks (DSC) provide an efficient solution to the problem of unsupervised subspace clustering by using an undercomplete deep auto-encoder with a fully-connected layer to exploit the self expressiveness property. This method uses undercomplete representations of the input data which makes it not so robust and more dependent on pre-training. To overcome this, we propose a simple yet efficient alternative method - Overcomplete Deep Subspace Clustering Networks (ODSC) where we use overcomplete representations for subspace clustering. In our proposed method, we fuse the features from both undercomplete and overcomplete auto-encoder networks before passing them through the self-expressive layer thus enabling us to extract a more meaningful and robust representation of the input data for clustering. Experimental results on four benchmark datasets show the effectiveness of the proposed method over DSC and other clustering methods in terms of clustering error. Our method is also not as dependent as DSC is on where pre-training should be stopped to get the best performance and is also more robust to noise.