WACV 2021 Papers

Unsupervised domain adaptation (UDA) deals with the adaptation process of a given source domain with labeled training data to a target domain for which only unannotated data is available. This is a challenging task as the domain shift leads to degraded performance on the target domain data if not addressed. In this paper, we analyze commonly used UDA classification datasets and discover systematic problems with regard to dataset setup, ground truth ambiguity and annotation quality. We manually clean the most popular UDA dataset in the research area (Office-31) and quantify the negative effects of inaccurate annotations through thorough experiments. Based on these insights, we collect the Adaptiope dataset - a large scale, diverse UDA dataset with synthetic, product and real world data - and show that its transfer tasks provide a challenge even when considering recent UDA algorithms. Our datasets are available at https://gitlab.com/tringwald/adaptiope.

We propose a Bayesian approximation to a deep learning architecture for 3D hand pose estimation. Through this framework, we explore and analyse the two types of uncertainties that are influenced either by data or by the learning capability. Furthermore, we draw comparisons against the standard estimator over three popular benchmarks. The first contribution lies in outperforming the baseline while in the second part we address the active learning application. We also show that with a newly proposed acquisition function, our Bayesian 3D hand pose estimator obtains lowest errors with the least amount of data. The underlying code is publicly available at https://github.com/razvancaramalau/al_bhpe.

Shape models have been used extensively to regularise segmentation of objects of interest in images, e.g. bones in medical x-ray radiographs, given supervised training examples. However, approaches usually adopt simple linear models that do not capture uncertainty and require extensive annotation effort to label a large number of set template landmarks for training. Conversely, supervised deep learning methods have been used on appearance directly (no explicit shape modelling) but these fail to capture detailed features that are clinically important. We present a supervised approach that combines both a non-linear generative shape model and a discriminative appearance-based convolutional neural network whilst quantifying uncertainty and relaxes the need for detailed, template based alignment for the training data. Our Bayesian framework couples the uncertainty from both the generator and the discriminator; our main contribution is the marginalisation of an intractable integral through the use of radial basis function approximations. We illustrate this model on the problem of segmenting bones from Psoriatic Arthritis hand radiographs and demonstrate that we can accurately measure the clinically important joint space gap between neighbouring bones.

This paper focuses on weakly-supervised action alignment, where only the ordered sequence of video-level actions is available for training. We propose a novel Duration Network, which captures a short temporal window of the video and learns to predict the remaining duration of a given action at any point in time with a level of granularity based on the type of that action. Further, we introduce a Segment-Level Beam Search to obtain the best alignment, that maximizes our posterior probability. Segment-Level Beam Search efficiently aligns actions by considering only a selected set of frames that have more confident predictions. The experimental results show that our alignments for long videos are more robust than existing models. Moreover, the proposed method achieves state of the art results in certain cases on the popular Breakfast and Hollywood Extended datasets.

Optimizing the discriminator in Generative Adversarial Networks (GANs) to completion in the inner training loop is computationally prohibitive, and on finite datasets would result in overfitting. To address this, a common update strategy is to alternate between k optimization steps for the discriminator D and one optimization step for the generator G. This strategy is repeated in various GAN algorithms where k is selected empirically. In this paper, we show that this update strategy is not optimal in terms of accuracy and convergence speed, and propose a new update strategy for networks with Wasserstein GAN (WGAN) group related loss functions (e.g. WGAN, WGAN-GP, Deblur GAN, and Super resolution GAN). The proposed update strategy is based on a loss change ratio comparison of G and D. We demonstrate that the proposed strategy improves both convergence speed and accuracy.

In this paper, we propose a novel approach for generalized zero-shot learning in a multi-modal setting, where we have novel classes of audio/video during testing that are not seen during training. We use the semantic relatedness of text embeddings as a means for zero-shot learning by aligning audio and video embeddings with the corresponding class label text feature space. Our approach uses a cross-modal decoder and a composite triplet loss. The cross-modal decoder enforces a constraint that the class label text features can be reconstructed from the audio and video embeddings of data points. This helps the audio and video embeddings to move closer to the class label text embedding. The composite triplet loss makes use of the audio, video, and text embeddings. It helps bring the embeddings from the same class closer and push away the embeddings from different classes in a multi-modal setting. This helps the network to perform better on the multi-modal zero-shot learning task. Importantly, our multi-modal zero-shot learning approach works even if a modality is missing at test time. We test our approach on the generalized zero-shot classification and retrieval tasks and show that our approach outperforms other models in the presence of a single modality as well as in the presence of multiple modalities. We validate our approach by comparing it with previous approaches and using various ablations.

Image matting is a fundamental computer vision problem and has many applications. Previous image matting methods always focus on extracting a general object or portrait from the background in an image. In this paper, we try to solve the text matting problem, which extracts characters (usually WordArts) from the background in an image. Different from traditional image matting problems, text matting is much harder because of its foreground's three properties: smallness, multi-objectness, and complicated structures and boundaries. We propose a two-stage attentional text matting pipeline to solve the text matting problem. In the first stage, we utilize text detection methods to serve as the attention mechanism. In the second stage, we employ the attentional text regions and matting system to obtain mattes of these text regions. Finally, we post-process the mattes and obtain the final matte of the input image. We also construct a large-scale dataset with high-quality annotations consisting of 46,289 unique foregrounds to facilitate the learning and evaluation of text matting. Extensive experiments on this dataset and real images clearly demonstrate the superiority of our proposed pipeline over previous image matting methods on the task of text matting.

Machine vision for precision agriculture has attracted considerable research interest in recent years. The goal of this paper is to develop an end-end cranberry health monitoring system to enable and support real time cranberry over-heating assessment to facilitate informed decisions that may sustain the economic viability of the farm. Toward this goal, we propose two main deep learning-based modules for: 1) cranberry fruit segmentation to delineate the exact fruit regions in the cranberry field image that are exposed to sun, 2) prediction of cloud coverage conditions to estimate the inner temperature of exposed cranberries We develop drone-based field data and ground-based sky data collection systems to collect video imagery at multiple time points for use in crop health analysis. Extensive evaluation on the data set shows that it is possible to predict exposed fruit's inner temperature with high accuracy (0.02% MAPE) when irradiance is predicted with 5.59-19.84% MAPE in the 5-20 minutes time horizon. With 62.54% mIoU for segmentation and 13.46 MAE for counting accuracies in exposed fruit identification, this system is capable of giving informed feedback to growers to take precautionary action (e.g., irrigation) in identified crop field regions with higher risk of sunburn in the near future. Though this novel system is applied for cranberry health monitoring, it represents a pioneering step forward in efficiency for farming and is useful in precision agriculture beyond the problem of cranberry overheating.

We deal with the problem of cross-task and cross-domain knowledge transfer in the realm of scene understanding for autonomous vehicles. We consider the scenario where supervision is available for a pair of tasks in a source domain while it is available for only one of the tasks in the target domain. Given that, the goal is to perform inference for the task in the target which is devoid of any training information. We argue that the only reported work in learning across tasks and domains (AT/DT) [23] faces the problem of domain shift between the source and target domains, hindering predictions on the target domain when the transfer of knowledge is learned on a statistically different yet related source domain. As a remedy, we develop a novel framework called ADA-AT/DT based on the adversarial training strategy to ensure that the domain-gaps are minimized for the common cross-domain supervised task. This, in effect, helps in realizing a domain-independent task-transfer function that eventually helps in performing improved inference in the target domain. We demonstrate that our proposed method significantly outperforms [23] by using models with 81% fewer trainable parameters. In addition, we perform experiments on a transformation mapping similar to U-Net to ensure maximum exploitation of features for task transfer. Extensive experiments have been performed on four different domains (Synthia, CityScapes, Carla, and KITTI) for two visual tasks (depth estimation and semantic segmentation) to confirm the superiority of our method.

Coronavirus Disease 2019 (COVID-19) has spread aggressively across the world causing an existential health crisis. Thus, having a system that automatically detects COVID-19 in tomography (CT) images can assist in quantifying the severity of the illness. Unfortunately, labelling chest CT scans requires significant domain expertise, time, and effort. We address these labelling challenges by only requiring point annotations, a single pixel for each infected region on a CT image. This labeling scheme allows annotators to label a pixel in a likely infected region, only taking 1-3 seconds, as opposed to 10-15 seconds to segment a region. Conventionally, segmentation models train on point-level annotations using the cross-entropy loss function on these labels. However, these models often suffer from low precision. Thus, we propose a consistency-based (CB) loss function that encourages the output predictions to be consistent with spatial transformations of the input images. The experiments on 3 open-source COVID-19 datasets show that this loss function yields significant improvement over conventional point-level loss functions and almost matches the performance of models trained with full supervision with much less human effort. Code is available at: https://github.com/IssamLaradji/covid19_weak_supervision.

This paper proposes to use the three vectors in a rotation matrix as the representation in head pose estimation and develops a new neural network based on the characteristic of such representation. We address two potential issues existed in current head pose estimation works: 1. Public datasets for head pose estimation use either Euler angles or quaternions to annotate data samples. However, both of these annotations have the issue of discontinuity and thus could result in some performance issues in neural network training. 2. Most research works report Mean Absolute Error (MAE) of Euler angles as the measurement of performance. We show that MAE may not reflect the actual behavior especially for the cases of profile views. To solve these two problems, we propose a new annotation method which uses three vectors to describe head poses and a new measurement Mean Absolute Error of Vectors (MAEV) to assess the performance. We also train a new neural network to predict the three vectors with the constraints of orthogonality. Our proposed method achieves state-of-the-art results on both AFLW2000 and BIWI datasets. Experiments show our vector-based annotation method can effectively reduce prediction errors for large pose angles.

In the last few years, deep neural networks' compression has become an important strand of machine learning and computer vision research. Deep models require sizeable computational complexity and storage when used, for instance, for Human Action Recognition (HAR) from videos, making them unsuitable to be deployed on edge devices. In this paper, we address this issue and propose a method to effectively compress Recurrent Neural Networks (RNNs) such as Gated Recurrent Units (GRUs) and Long-Short-Term-Memory Units (LSTMs) that are used for HAR. We use a Variational Information Bottleneck (VIB) theory-based pruning approach to limit the information flow through the sequential cells of RNNs to a small subset. Further, we combine our pruning method with a specific group-lasso regularization technique that significantly improves compression. The proposed techniques reduce model parameters and memory footprint from latent representations, with little or no reduction in the validation accuracy while increasing the inference speed several-fold. We perform experiments on the three widely used Action Recognition datasets, viz. UCF11, HMDB51, and UCF101, to validate our approach. We show that our method achieves over 70 times greater compression than the nearest competitor with comparable accuracy for action recognition on UCF11.

Several coded exposure techniques have been proposed for acquiring high frame rate videos at low bandwidth. Most recently, a Coded-2-Bucket camera has been proposed that can acquire two compressed measurements in a single exposure, unlike previously proposed coded exposure techniques, which can acquire only a single measurement. Although two measurements are better than one for an effective video recovery, we are yet unaware of the clear advantage of two measurements, either quantitatively or qualitatively. Here, we propose a unified learning-based framework to make such a qualitative and quantitative comparison between those which capture only a single coded image (Flutter Shutter, Pixel-wise coded exposure) and those that capture two measurements per exposure (C2B). Our learning-based framework consists of a shift-variant convolutional layer followed by a fully convolutional deep neural network. Our proposed unified framework achieves the state of the art reconstructions in all three sensing techniques. Further analysis shows that when most scene points are static, the C2B sensor has a significant advantage over acquiring a single pixel-wise coded measurement. However, when most scene points undergo motion, the C2B sensor has only a marginal benefit over the single pixel-wise coded exposure measurement.

We propose a novel framework to register sports-fields as they appear in broadcast sports videos. Unlike previous approaches, we particularly address the challenge of field-registration when: (a) there are not enough distinguishable features on the field, and (b) no prior knowledge is available about the camera. To this end, we detect a grid of keypoints distributed uniformly on the entire field instead of using only sparse local corners and line intersections, thereby extending the keypoint coverage to the texture-less parts of the field as well. To further improve keypoint based homography estimate, we differentialbly warp and align it with a set of dense field-features defined as normalized distance-map of pixels to their nearest lines and key-regions. We predict the keypoints and dense field-features simultaneously using a multi-task deep network to achieve computational efficiency. To have a comprehensive evaluation, we have compiled a new dataset called SportsFields which is collected from 192 video-clips from 5 different sports covering large environmental and camera variations. We empirically demonstrate that our algorithm not only achieves state of the art field-registration accuracy but also runs in real-time for HD resolution videos using commodity hardware.

In this paper, we present a novel, end-to-end 6D object pose estimation method that operates on RGB inputs. Our approach is composed of 2 main components: the first component classifies the objects in the input image and proposes an initial 6D pose estimate through a multi-task, CNN-based encoder/multi-decoder module. The second component, a refinement module, includes a renderer and a multi-attentional pose refinement network, which iteratively refines the estimated poses by utilizing both appearance features and flow vectors. Our refiner takes advantage of the hybrid representation of the initial pose estimates to predict the relative errors with respect to the target poses. It is further augmented by a spatial multi-attention block that emphasizes objects' discriminative feature parts. Experiments on three benchmarks for 6D pose estimation show that our proposed pipeline outperforms state-of-the-art RGB-based methods with competitive runtime performance.

We propose a new approach to Human Activity Evaluation (HAE) in long videos using graph-based multi-task modeling. Previous works in activity evaluation either directly compute a metric using a detected skeleton or use the scene information to regress the activity score. These approaches are insufficient for accurate activity assessment since they only compute an average score over a clip, and do not consider the correlation between the joints and body dynamics. Moreover, they are highly scene-dependent which makes the generalizability of these methods questionable. We propose a novel multi-task framework for HAE that utilizes a Graph Convolutional Network backbone to embed the interconnections between human joints in the features. In this framework, we solve the Human Activity Segmentation (HAS) problem as an auxiliary task to improve activity assessment. The HAS head is powered by an encoder-Decoder Temporal Convolutional Network to semantically segment long videos into distinct activity classes, whereas, HAE uses a Long-Short-Term-Memory-based architecture. We evaluate our method on the UW-IOM and TUM Kitchen datasets and discuss the success and failure cases in these two datasets.

We propose a new algorithm to incorporate class conditional information into the critic of GANs via a multi-class generalization of the commonly used Hinge loss that is compatible with both supervised and semi-supervised settings. We study the compromise between training a state of the art generator and an accurate classifier simultaneously, and propose a way to use our algorithm to measure the degree to which a generator and critic are class conditional. We show the trade-off between a generator-critic pair respecting class conditioning inputs and generating the highest quality images. With our multi-hinge loss modification we are able to improve Inception Scores and Frechet Inception Distance on the Imagenet dataset.

Rotoscoping of facial features is often an integral part of Visual Effects post-production, where the parametric contours created by artists need to be highly detailed, consist of multiple interacting components, and involve significant manual supervision. Yet those assets are usually discarded after compositing and hardly reused. In this paper, we present the first methodology to learn from these assets. With only a few manually rotoscoped shots, we identify and extract semantically consistent and task specific landmark points and re-vectorize the roto shapes based on these landmarks. We then train two separate models -- one to predict landmarks based on a rough crop of the face region, and the other to predict the roto shapes using only the inferred landmarks from the first model. In preliminary production testing, 26% of shots rotoscoped using our tool were able to be used with no adjustment, and another 47% were able to be used with minor adjustments. This represents a significant time savings for the studio, as artists are able to rotoscope almost 73% of their shots with no manual rotoscoping and some spline adjustment. This paper presents a novel application of machine learning to professional interactive rotoscoping, a methodology to convert unstructured roto shapes into a self-annotated, trainable dataset that can be harnessed to make accurate predictions on future shots of a similar object, and a limited dataset of rotoscoped multi-shape fine feature systems from a real film production.

Thermal face imagery, which captures the naturally emitted heat from the face, is limited in availability compared to face imagery in the visible spectrum. To help address this scarcity of thermal face imagery for research and algorithm development, we present the DEVCOM Army Research Laboratory Visible-Thermal Face Dataset (ARL-VTF). With over 500,000 images from 395 subjects, the ARL-VTF dataset represents, to the best of our knowledge, the largest collection of paired visible and thermal face images to date. The data was captured using a modern long wave infrared (LWIR) camera mounted alongside a stereo setup of three visible spectrum cameras. Variability in expressions, pose, and eyewear has been systematically recorded. The dataset has been curated with extensive annotations, metadata, and standardized protocols for evaluation. Furthermore, this paper presents extensive benchmark results and analysis on thermal face landmark detection and thermal-to-visible face verification by evaluating state-of-the-art models on the ARL-VTF dataset.