WACV 2021 Papers

Oriented object detection in aerial images is a challenging task as the objects in aerial images are displayed in arbitrary directions and are usually densely packed. Current oriented object detection methods mainly rely on two-stage anchor-based detectors. However, the anchor-based detectors typically suffer from a severe imbalance issue between the positive and negative anchor boxes. To address this issue, in this work we extend the horizontal keypoint-based object detector to the oriented object detection task. In particular, we first detect the center keypoints of the objects, based on which we then regress the box boundary-aware vectors (BBAVectors) to capture the oriented bounding boxes. The box boundary-aware vectors are distributed in the four quadrants of a Cartesian coordinate system for all arbitrarily oriented objects. To relieve the difficulty of learning the vectors in the corner cases, we further classify the oriented bounding boxes into horizontal and rotational bounding boxes. In the experiment, we show that learning the box boundary-aware vectors is superior to directly predicting the width, height, and angle of an oriented bounding box, as adopted in the baseline method. Besides, the proposed method competes favorably with state-of-the-art methods. Code is available at https://github.com/yijingru/BBAVectors-Oriented-Object-Detection.

We humans can impeccably search for a target object, given its name only, even in an unseen environment. We argue this ability is largely due to three main reasons: the incorporation of prior knowledge (or experience), the adaptation of it to the new environment using the observed visual cues and most importantly optimistically searching without giving up early.This is currently missing in the state-of-the-art visual navigation methods based on Reinforcement Learning (RL). In this paper, we propose to use externally learned prior knowledge of the relative object locations and integrate it into our model by constructing a neural graph. In order to efficiently incorporate the graph without increasing the state-space complexity, we propose Graph-based Value Estimation (GVE) module. GVE provides a more accurate baseline for estimating the Advantage function in actor-critic RL algorithm. This results in reduced value estimation error and, consequently, convergence to a more optimal policy. Through empirical studies, we show that our agent, dubbed as the optimistic agent, has a more realistic estimate of the state value during a navigation episode which leads to a higher success rate. Our extensive ablation studies show the efficacy of our simple method which achieves the state-of-the-art results measured by the conventional visual navigation metrics, e.g. Success Rate (SR) and Success weighted by Path Length (SPL), in AI2THOR environment.

Event classification can add valuable information for semantic search and the increasingly important topic of fact validation in news. So far, only few approaches address image classification for newsworthy event types such as natural disasters, sports events, or elections. Previous work distinguishes only between a limited number of event types and relies on rather small datasets for training. In this paper, we present a novel ontology-driven approach for the classification of event types in images. We leverage a large number of real-world news events to pursue two objectives: First, we create an ontology based on Wikidata comprising the majority of event types. Second, we introduce a novel large-scale dataset that was acquired through Web crawling. Several baselines are proposed including an ontology-driven learning approach that aims to exploit structured information of a knowledge graph to learn relevant event relations using deep neural networks. Experimental results on existing as well as novel benchmark datasets demonstrate the superiority of the proposed ontology-driven approach.

Understanding temporal information and how the visual world changes over time, is a fundamental ability of intelligent systems. In video understanding, temporal information is at the core of many current challenges, including compression, efficient inference, motion estimation or summarization. However, in current video datasets it has been observed that action classes can often be recognized without any temporal information, from a single frame of video. As a result, both benchmarking and training in these datasets may give an unintentional advantage to models with strong image understanding capabilities, as opposed to those with strong temporal understanding, potentially hindering progress. In this paper we address this problem head on by identifying action classes where temporal information is actually necessary to recognize them and call these "temporal classes". Selecting temporal classes using a computational method would bias the process. Instead, we propose a methodology based on a simple and effective human annotation experiment. We remove just the temporal information, by shuffling frames in time, and measure if the action can still be recognized. Classes that cannot be recognized when frames are not in order, are included in the temporal set. We observe that this set is statistically different from other static classes, and that performance in it correlates with a network's ability to capture temporal information. Thus we use it as a benchmark on current popular networks, which reveals a series of interesting facts, like inflated convolutions bias networks towards classes where motion is not important. We also explore the effect of training on the temporal set, and observe that this leads to better generalization in unseen classes, demonstrating the need for more temporal data. We hope that the proposed dataset of temporal categories will help guide future research in temporal modeling for better video understanding.

Humans have the innate ability to recognize new objects just by looking at sketches of them (also referred as to prototype images). Similarly, prototypical images can be used as an effective visual representations of unseen classes to tackle few-shot learning (FSL) tasks. Our main goal is to recognize unseen hand signs (gestures) traffic-signs, and corporate-logos, by having their iconographic images or prototypes. Previous works proposed to utilize variational prototypical-encoders (VPE) to address FSL problems. While VPE learns an image-to-image translation task efficiently, we discovered that its performance is significantly hampered by the so-called hubness problem and it fails to regulate the representations in the latent space. Hence, we propose a new model (VPE++) that inherently reduces hubness and incorporates contrastive and multi-task losses to increase the discriminative ability of FSL models. Results show that the VPE++ approach can generalize better to the unseen classes and can achieve superior accuracies on logos, traffic signs, and hand gestures datasets as compared to the state-of-the-art.

Despite the initial belief that Convolutional Neural Networks (CNNs) are driven by shapes to perform visual recognition tasks, recent evidence suggests that texture bias in CNNs provides higher performing and more robust models. This contrasts with the perceptual bias in the human visual cortex, which has a stronger preference towards shape components. Perceptual differences may explain why CNNs achieve human-level performance when large labeled datasets are available, but their performance significantly degrades in low-labeled data scenarios, such as few-shot semantic segmentation. To remove the texture bias in the context of few-shot learning, we propose a novel architecture that integrates a set of Difference of Gaussians (DoG) to attenuate high-frequency local components in the feature space. This produces a set of modified feature maps, whose high-frequency components are diminished at different standard deviation values of the Gaussian distribution in the spatial domain. As this results in multiple feature maps for a single image, we employ a bi-directional convolutional long-short-term-memory to efficiently merge the multi scale-space representations. We perform extensive experiments on three well-known few-shot segmentation benchmarks --Pascal i5, COCO-20i and FSS-1000-- and demonstrate that our method outperforms state-of-the-art approaches in two datasets under the same conditions.

State-of-the-art learning-based monocular 3D reconstruction methods learn priors over object categories on the training set, and as a result struggle to achieve reasonable generalization to object categories unseen during training. In this paper we study the inductive biases encoded in the model architecture that impact the generalization of learning-based 3D reconstruction methods. We find that 3 inductive biases impact performance: the spatial extent of the encoder, the use of the underlying geometry of the scene to describe point features, and the mechanism to aggregate information from multiple views. Additionally, we propose mechanisms to enforce those inductive biases: a point representation that is aware of camera position, and a variance cost to aggregate information across views. Our model achieves state-of-the-art results on the standard ShapeNet 3D reconstruction benchmark in various settings.

Object recognition ability is indispensable for robots to act like humans in a home environment. For example, when considering an object searching task, humans can recognize a naturally arranged object previously held in their hands while ignoring never observed objects. Even in such a simple task, we need to deal with three complex problems: domain adaptation, open-set recognition, and continual learning. However, most existing datasets are simplified to focus on one problem and do not measure the object recognition ability for home robots when multiple problems are simultaneously present. In this paper, we propose the COSDA-HR (Continual Open Set Domain Adaptation for Home Robot) dataset that requires dealing with the above three problems simultaneously. The COSDA-HR dataset focuses particularly on the scenario in which naturally arranged objects in a room are recognized by training with handheld objects towards the goal of creating a user-friendly teaching system for home robots. We provide various baselines to address the problems in the COSDA-HR dataset by combining state-of-the-art methods from each research area and analyze the limitations of such simple combinations. We consider that it is necessary to study the methods of handling multiple problems simultaneously instead of solving each problem to realize practical object recognition systems for home robots.

We introduce a principled approach for synthesizing new views of a scene given a single source image. Previous methods for novel view synthesis can be divided into image-based rendering methods (e.g, flow prediction) or pixel generation methods. Flow predictions enable the target view to re-use pixels directly, but can easily lead to distorted results. Directly regressing pixels can produce structurally consistent results but generally suffer from the lack of low-level details. In this paper, we utilize an encoder-decoder architecture to regress pixels of a target view. In order to maintain details, we couple the decoder aligned feature maps with skip connections, where the alignment is guided by predicted depth map of the target view. Our experimental results show that our method does not suffer from distortions and successfully preserves texture details with aligned skip connections.

Deep neural networks (DNNs) fail to learn effectively under label noise and have been shown to memorize random labels which affect their generalization performance. We consider learning in isolation, using one-hot encoded labels as the sole source of supervision, and a lack of regularization to discourage memorization as the major shortcomings of the standard training procedure. Thus, we propose Noisy Concurrent Training (NCT) which leverages collaborative learning to use the consensus between two models as an additional source of supervision. Furthermore, inspired by trial-to-trial variability in the brain, we propose a counterintuitive regularization technique, target variability, which entails randomly changing the labels of a percentage of training samples in each batch as a deterrent to memorization and overgeneralization in DNNs. Target variability is applied independently to each model to keep them diverged and avoid the confirmation bias. As DNNs tend to prioritize learning simple patterns first before memorizing the noisy labels, we employ a dynamic learning scheme whereby as the training progresses, the two models increasingly rely more on their consensus. NCT also progressively increases the target variability to avoid memorization in later stages. We demonstrate the effectiveness of our approach on both synthetic and real-world noisy benchmark datasets.

While noise is commonly considered a nuisance in computing systems, a number of studies in neuroscience have shown several benefits of noise in the nervous system from enabling the brain to carry out computations such as probabilistic inference as well as carrying additional information about the stimuli. Similarly, noise has been shown to improve the performance of deep neural networks. In this study, we further investigate the effect of adding noise in the knowledge distillation framework because of its resemblance to collaborative subnetworks in the brain regions. We empirically show that injecting constructive noise at different levels in the collaborative learning framework enables us to train the model effectively and distill desirable characteristics in the student model. In doing so, we propose three different methods that target the common challenges in deep neural networks: minimizing the performance gap between a compact model and large model (Fickle Teacher), training high performance compact adversarially robust models (Soft Randomization), and training models efficiently under label noise (Messy Collaboration). Our findings motivate further study in the role of noise as a resource for learning in a collaborative learning framework.

We propose methods for estimating correspondence between two point sets under the presence of outliers in both the source and target sets. The proposed algorithms expand upon the theory of the regression without correspondence problem to estimate transformation coefficients using unordered multisets of covariates and responses. Previous theoretical analysis of the problem has been done in a setting where the responses are a complete permutation of the regressed covariates. This paper expands the problem setting by analyzing the cases where only a subset of the responses is a permutation of the regressed covariates in addition to some covariates possibly being adversarial outliers. We term this problem robust regression without correspondence and provide several algorithms based on random sample consensus for exact and approximate recovery in a noiseless and noisy one-dimensional setting as well as an approximation algorithm for multiple dimensions. The theoretical guarantees of the algorithms are verified in simulated data. We demonstrate an important computational neuroscience application of the proposed framework by demonstrating its effectiveness in a Caenorhabditis elegans neuron matching problem where the presence of outliers in both the source and target nematodes is a natural tendency.

Contrast materials are often injected into body to contrast specific tissues in Computed Tomography (CT) images. Contrast Enhanced CT (CECT) images obtained in this way are more useful than Non-Enhanced CT (NECT) images for medical diagnosis, but not available for everyone due to side effects of the contrast materials. Motivated by this, we develop a neural network that takes NECT images and generates their CECT counterparts. Learning such a network is extremely challenging since NECT and CECT images for training are not aligned even at the same location of the same patient due to movements of internal organs. We propose a two-stage framework to address this issue. The first stage trains an auxiliary network that removes the effect of contrast enhancement in CECT images to synthesize their NECT counterparts well-aligned with them. In the second stage, the target model is trained to predict the real CECTimages given a synthetic NECT image as input. Experimental results and analysis by physicians on abdomen CT images suggest that our method outperforms existing models for neural image synthesis.

In this paper, we present a novel differential morph detection framework, utilizing landmark and appearance disentanglement. In our framework, the face image is represented in the embedding domain using two disentangled but complementary representations. The network is trained by triplets of face images, in which the intermediate image inherits the landmarks from one image and the appearance from the other image. This initially trained network is further trained for each dataset using contrastive representations. We demonstrate that, by employing appearance and landmark disentanglement, the proposed framework can provide state-of-the-art differential morph detection performance. This functionality is achieved by the using distances in landmark, appearance, and ID domains. The performance of the proposed framework is evaluated using three morph datasets generated with different methodologies.

Predicting the future trajectories of surrounding vehicles is a key competence for safe and efficient real-world autonomous driving systems. Previous works have presented deep neural network models for predictions using a detailed prior map which includes driving lanes and explicitly expresses the road rules like legal traffic directions and valid paths through intersections. Since it is unrealistic to assume the existence of the detailed prior maps for all areas, we use a map generated from only perceptual data (3D points measured by a LiDAR sensor). Such maps do not explicitly denote road rules, which makes prediction tasks more difficult. To overcome this problem, we propose a novel generative adversarial network (GAN) based framework. A discriminator in our framework can distinguish whether predicted trajectories follow road rules, and a generator can predict trajectories following it. Our framework implicitly extracts road rules by projecting trajectories onto the map via a differentiable function and training positional relations between trajectories and obstacles on the map. We also extend our framework to multimodal predictions so that various future trajectories are predicted. Experimental results show that our method outperforms other state-of-the-art methods in terms of trajectory errors and the ratio of trajectories that fall on drivable lanes.

Although providing exceptional results for many computer vision tasks, state-of-the-art deep learning algorithms catastrophically struggle in low data scenarios. However, if data in additional modalities exist (e.g. text) this can compensate for the lack of data and improve the classification results. To overcome this data scarcity, we design a cross-modal feature generation framework capable of enriching the low populated embedding space in few-shot scenarios, leveraging data from the auxiliary modality. Specifically, we train a generative model that maps text data into the visual feature space to obtain more reliable prototypes. This allows to exploit data from additional modalities (e.g. text) during training while the ultimate task at test time remains classification with exclusively visual data. We show that in such cases nearest neighbor classification is a viable approach and outperform state-of-the-art single-modal and multimodal few-shot learning methods on the CUB-200 and Oxford-102 datasets.

A great number of situational comedies (sitcoms) are being regularly made and the task of adding laughter tracks to these is a critical task. Providing an ability to be able to predict whether something will be humorous to the audience is also crucial. In this project, we aim to automate this task. Towards doing so, we annotate an existing sitcom (`Big Bang Theory') and use the laughter cues present to obtain a manual annotation for this show. We provide detailed analysis for the dataset design and further evaluate various state of the art baselines for solving this task. We observe that existing LSTM and BERT based networks on the text alone do not perform as well as joint text and video or only video-based networks. Moreover, it is challenging to ascertain that the words attended to while predicting laughter are indeed humorous. Our dataset and analysis provided through this paper is a valuable resource towards solving this interesting semantic and practical task. As an additional contribution, we have developed a novel model for solving this task that is a multi-modal self-attention based model that outperforms currently prevalent models for solving this task. The project page for the submission is \url https://multimodal-humor-dataset.github.io/

While accurate disease prediction from retinal fundus images is critical, collecting large amounts of high quality labeled training data to build such supervised models is difficult. Deep learning classifiers have led to high accuracy results across a wide variety of medical imaging problems, but they need large amounts of labeled data. Given a fundus image, we aim to evaluate various solutions for learning deep neural classifiers using small labeled data for three tasks related to eye disease prediction: (T1) predicting one of the five broad categories - diabetic retinopathy, age-related macular degeneration, glaucoma, melanoma and normal, (T2) predicting one of the 320 fine-grained disease sub-categories, (T3) generating a textual diagnosis. The problem is challenging because of small data size, need for predictions across multiple tasks, handling image variations, and large number of hyper-parameter choices. Modeling the problem under a multi-task learning (MTL) setup, we investigate the contributions of each of the proposed tasks while dealing with a small amount of labeled data. Further, we suggest a novel MTL-based teacher ensemble method for knowledge distillation. On a dataset of 7212 labeled and 35854 unlabeled images across 3502 patients, our technique obtains 83% accuracy, 75% top-5 accuracy and 48 BLEU for tasks T1, T2 and T3 respectively. Even with 15% training data, our method outperforms baselines by 8.1, 3.2 and 11.2 points for the three tasks respectively.

Learning multiple domains/tasks with a single model is important for improving data efficiency and lowering inference cost for numerous vision tasks, especially on resource-constrained mobile devices. However, hand-crafting a multi-domain/task model can be both tedious and challenging. This paper proposes a novel approach to automatically learn a multi-path network for multi-domain visual classification on mobile devices. The proposed multi-path network is learned from neural architecture search by applying one reinforcement learning controller for each domain to select the best path in the super-network created from a MobileNetV3-like search space. An adaptive balanced domain prioritization algorithm is proposed to balance optimizing the joint model on multiple domains simultaneously. The determined multi-path model selectively shares parameters across domains in shared nodes while keeping domain-specific parameters within non-shared nodes in individual domain paths. This approach effectively reduces the total number of parameters and FLOPS, encouraging positive knowledge transfer while mitigating negative interference across domains. Extensive evaluations on the Visual Decathlon dataset demonstrate that the proposed multi-path model achieves state-of-the-art performance in terms of accuracy, model size, and FLOPS against other approaches using MobileNetV3-like architectures. Furthermore, the proposed method improves average accuracy over learning single-domain models individually, and reduces the total number of parameters and FLOPS by 78% and 32% respectively, compared to the approach that simply bundles single-domain models for multi-domain learning.