WACV 2021 论文列表

Scene text instances found in natural images carry explicit semantic information that can provide important cues to solve a wide array of computer vision problems. In this paper, we focus on leveraging multi-modal content in the form of visual and textual cues to tackle the task of fine-grained image classification and retrieval. First, we obtain the text instances from images by employing a text reading system. Then, we combine textual features with salient image regions to exploit the complementary information carried by the two sources. Specifically, we employ a Graph Convolutional Network to perform multi-modal reasoning and obtain relationship-enhanced features by learning a common semantic space between salient objects and text found in an image. By obtaining an enhanced set of visual and textual features, the proposed model greatly outperforms previous state-of-the-art in two different tasks, fine-grained classification and image retrieval in the Con-Text and Drink Bottle datasets.

Many interesting tasks in machine learning and computer vision are learned by optimising an objective function defined as a weighted linear combination of multiple losses. The final performance is sensitive to choosing the correct (relative) weights for these losses. Finding a good set of weights is often done by adopting them into the set of hyper-parameters, which are set using an extensive grid search. This is computationally expensive. In this paper, we propose a weighting scheme based on the coefficient of variations and set the weights based on properties observed while training the model. The proposed method incorporates a measure of uncertainty to balance the losses, and as a result the loss weights evolve during training without requiring another (learning based) optimisation. In contrast to many loss weighting methods in literature, we focus on single-task multi-loss problems, such as monocular depth estimation and semantic segmentation, and show that multi-task approaches for loss weighting do not work on those single-tasks. The validity of the approach is shown empirically for depth estimation and semantic segmentation on multiple datasets.

This paper presents a novel multi-level generative chaotic Recurrent Neural Network (RNN) for image inpainting. This technique utilizes a general framework with multiple chaotic RNN that makes learning the image prior from a single corrupted image more robust and efficient. The proposed network utilizes a randomly-initialized process for parameterization, along with a unique quad-directional encoder structure, chaotic state transition, and adaptive importance for multi-level RNN updating. The efficacy of the approach has been validated through multiple experiments. In spite of a much lower computational load, quantitative comparisons reveal that the proposed approach exceeds the performance of several image restoration benchmarks.

Moving object segmentation (MOS) in different practical scenarios like weather degraded, dynamic background, etc. videos is a challenging and high demanding task for various computer vision applications. Existing supervised approaches achieve remarkable performance with complicated training or extensive fine-tuning or inappropriate training-testing data distribution. Also, the generalized effect of existing works with completely unseen data is difficult to identify. In this work, the recurrent feature sharing based generative adversarial network is proposed with unseen video analysis. The proposed network comprises of dilated convolution to extract the spatial features at multiple scales. Along with the temporally sampled multiple frames, previous frame output is considered as input to the network. As the motion is very minute between the two consecutive frames, the previous frame decoder features are shared with encoder features recurrently for current frame foreground segmentation. This recurrent feature sharing of different layers helps the encoder network to learn the hierarchical interactions between the motion and appearance based features. Also, the learning of the proposed network is concentrated in different ways, like disjoint and global training-testing for MOS. An extensive experimental analysis of the proposed network is carried out on two benchmark video datasets with seen and unseen MOS video. Qualitative and quantitative experimental study shows that the proposed network outperforms the existing methods.

Semantic segmentation of 3D point cloud data is essential for enhanced high-level perception in autonomous platforms. Furthermore, given the increasing deployment of LiDAR sensors onboard of cars and drones, a special emphasis is also placed on non-computationally intensive algorithms that operate on mobile GPUs. Previous efficient state-of-the-art methods relied on 2D spherical projection of point clouds as input for 2D fully convolutional neural networks to balance the accuracy-speed trade-off. This paper introduces a novel approach for 3D point cloud semantic segmentation that exploits multiple projections of the point cloud to mitigate the loss of information inherent in single projection methods. Our Multi-Projection Fusion (MPF) framework analyzes spherical and bird's-eye view projections using two separate highly-efficient 2D fully convolutional models then combines the segmentation results of both views. The proposed framework is validated on the SemanticKITTI dataset where it achieved a mIoU of 55.5 which is higher than state-of-the-art projection-based methods RangeNet++ [23] and PolarNet [44] while being 1.6x faster than the former and 3.1x faster than the latter.

Single-photon avalanche diodes (SPADs) are a rapidly developing image sensing technology with extreme low-light sensitivity and picosecond timing resolution. These unique capabilities have enabled SPADs to be used in applications like LiDAR, non-line-of-sight imaging and fluorescence microscopy that require imaging in photon-starved scenarios. In this work we harness these capabilities for dealing with motion blur in a passive imaging setting in low illumination conditions. Our key insight is that the data captured by a SPAD array camera can be represented as a 3D spatio-temporal tensor of photon detection events which can be integrated along arbitrary spatio-temporal trajectories with dynamically varying integration windows, depending on scene motion. We propose an algorithm that estimates pixel motion from photon timestamp data and dynamically adapts the integration windows to minimize motion blur. Our simulation results show the applicability of this algorithm to a variety of motion profiles including translation, rotation and local object motion. We also demonstrate the real-world feasibility of our method on data captured using a 32x32 SPAD camera.

Motorcycles are often related to transit and criminal issues due to its abundance in the transit. Despite its importance, motorcycles are a seldom addressed problem in the computer vision community. We credit this problem to the lack of large-scale datasets and strong baseline models. Therefore, we present the first large-scale Motorcycles Re-Identification (MoRe) dataset. MoRe consists of 3,827 individuals (i.e., the set of motorbikes and motorcyclist) captured by ten surveillance cameras placed in Brazil's urban traffic scenarios. Furthermore, we evaluate a deep learning model trained using well-known training tricks from the object re-identification literature to present a strong baseline for the motorcycle re-identification (ReID) problem. More importantly, we highlight some crucial problems in this topic as the influence of distractors and the domain shift. Experimental results demonstrate the effectiveness of the strong baseline model with an increase of at least 19.27 p.p. in the rank-1 when compared to the state-of-the-art in the BPReID dataset. Finally, we present some insights regarding the information learned by the strong baseline model when computing the similarities between motorcycle images.

In this paper, we propose risk-calibrated evidential deep classifiers to reduce the costs associated with classification errors. We use two main approaches. The first is to develop methods to quantify the uncertainty of a classifier's predictions and reduce the likelihood of acting on erroneous predictions. The second is a novel way to train the classifier such that erroneous classifications are biased towards less risky categories. We combine these two approaches in a principled way. While doing this, we extend evidential deep learning with pignistic probabilities, which are used to quantify uncertainty of classification predictions and model rational decision making under uncertainty. We evaluate the performance of our approach on several image classification tasks. We demonstrate that our approach allows to (i) incorporate misclassification cost while training deep classifiers, (ii) accurately quantify the uncertainty of classification predictions, and (iii) simultaneously learn how to make classification decisions to minimize expected cost of classification errors.

The paper presents a learning-based method for computing a discriminative 3D point cloud descriptor for place recognition purposes. Existing methods, such as PointNetVLAD, are based on unordered point cloud representation. They use PointNet as the first processing step to extract local features, which are later aggregated into a global descriptor. The PointNet architecture is not well suited to capture local geometric structures. Thus, state-of-the-art methods enhance vanilla PointNet architecture by adding different mechanism to capture local contextual information, such as graph convolutional networks or using hand-crafted features. We present an alternative approach, dubbed MinkLoc3D, to compute a discriminative 3D point cloud descriptor, based on a sparse voxelized point cloud representation and sparse 3D convolutions. The proposed method has a simple and efficient architecture. Evaluation on standard benchmarks proves that MinkLoc3D outperforms current state-of-the-art. Our code is publicly available on the project website.

This paper proposes minimal solvers that use combinations of imaged translational symmetries and parallel scene lines to jointly estimate lens undistortion with either affine rectification or focal length and absolute orientation. We use constraints provided by orthogonal scene planes to recover the focal length. We show that solvers using feature combinations can recover more accurate calibrations than solvers using only one feature type on scenes that have a balance of lines and texture. We also show that the proposed solvers are complementary and can be used together in a RANSAC-based estimator to improve auto-calibration accuracy. State-of-the-art performance is demonstrated on a standard dataset of lens-distorted urban images. The code is available at https://github.com/ylochman/single-view-autocalib.

Neural Module Network (NMN) exhibits strong interpretability and compositionality thanks to its handcrafted neural modules with explicit multi-hop reasoning capability. However, most NMNs suffer from two critical drawbacks: 1) scalability: customized module for specific function renders it impractical when scaling up to a larger set of functions in complex tasks; 2) generalizability: rigid pre-defined module inventory makes it difficult to generalize to unseen functions in new tasks/domains. To design a more powerful NMN architecture for practical use, we propose Meta Module Network (MMN) centered on a novel meta module, which can take in function recipes and morph into diverse instance modules dynamically. The instance modules are then woven into an execution graph for complex visual reasoning, inheriting the strong explainability and compositionality of NMN. With such a flexible instantiation mechanism, the parameters of instance modules are inherited from the central meta module, retaining the same model complexity as the function set grows, which promises better scalability. Meanwhile, as functions are encoded into the embedding space, unseen functions can be readily represented based on its structural similarity with previously observed ones, which ensures better generalizability. Experiments on GQA and CLEVR datasets validate the superiority of MMN over state-of-the-art NMN designs. Synthetic experiments on held-out unseen functions from GQA dataset also demonstrate the strong generalizability of MMN.

Binary Neural Networks (BNNs) are neural networks which use binary weights and activations instead of the typical 32-bit floating point values. They have reduced model sizes and allow for efficient inference on mobile or embedded devices with limited power and computational resources. However, the binarization of weights and activations leads to feature maps of lower quality and lower capacity and thus a drop in accuracy compared to their 32-bit counterparts. Previous work has increased the number of channels or used multiple binary bases to alleviate these problems. In this paper, we instead present an architectural approach: MeliusNet. It consists of alternating a DenseBlock, which increases the feature capacity, and our proposed ImprovementBlock, which increases the feature quality. Experiments on the ImageNet dataset demonstrate the superior performance of our MeliusNet over a variety of popular binary architectures with regards to both computation savings and accuracy. Furthermore, BNN models trained with our method can match the accuracy of the popular compact network MobileNet-v1 in terms of model size and number of operations.

In this work we present a novel approach for Unsupervised Video Object Segmentation, that is automatically generating instance level segmentation masks for salient objects and tracking them in a video. We efficiently handle problems present in existing methods such as drift while temporal propagation, tracking and addition of new objects. To this end, we propose a novel idea of improving masks in an online manner using ensemble of criteria whose task is to inspect the quality of masks. We introduce a novel idea of assessing mask quality using a neural network called Selector Net. The proposed network is trained is such way that it is generalizes across various datasets. Our proposed method is able to limit the noise accumulated along the video, giving state of the art result on Davis 2019 Unsupervised challenge dataset with J&F mean 61.6%. We also tested on datasets such as FBMS and SegTrack V2 and performed better or on par compared to the other methods.

DensePose estimation task is a significant step forward for enhancing user experience computer vision applications ranging from augmented reality to cloth fitting. Existing neural network models capable of solving this task are heavily parameterized and a long way from being transferred to an embedded or mobile device. To enable Dense Pose inference on the end device with current models, one needs to support an expensive server-side infrastructure and have a stable internet connection. To make things worse, mobile and embedded devices do not always have a powerful GPU inside. In this work, we target the problem of redesigning the DensePose R-CNN model's architecture so that the final network retains most of its accuracy but becomes more light-weight and fast. To achieve that, we tested and incorporated many deep learning innovations from recent years, specifically performing an ablation study on 23 efficient backbone architectures, multiple two-stage detection pipeline modifications, and custom model quantization methods. As a result, we achieved 17 times model size reduction and 2 times latency improvement compared to the baseline model.

Estimating 3D hand poses from a single RGB image is challenging because depth ambiguity leads the problem ill-posed. Training hand pose estimators with 3D hand mesh annotations and multi-view images often results in significant performance gains. However, existing multi-view datasets are relatively small with hand joints annotated by off-the-shelf trackers or automated through model predictions, both which may be inaccurate and can introduce biases. Collecting a large-scale multi-view 3D hand pose images with accurate mesh and joint annotations is valuable but strenuous. In this paper, we design a spin match algorithm that enables rigid mesh model matching without any target mesh ground truth. Based on the match algorithm, we propose an efficient pipeline to generate a large-scale multi-view hand mesh (MVHM) dataset with accurate 3D hand mesh and joint labels. We further present a multi-view hand pose estimation approach to verify that training a hand pose estimator with our generated dataset greatly enhances the performance. Experimental results show that our approach achieves the performance of 0.990 in \text AUC _ \text 20-50 on the MHP dataset compared to the previous state-of-the-art of 0.939 on this dataset.

Deep neural networks are powerful, massively parameterized machine learning models that have been shown to perform well in supervised learning tasks. However, very large amounts of labeled data are usually needed to train deep neural networks. Several semi-supervised learning approaches have been proposed to train neural networks using smaller amounts of labeled data with a large amount of unlabeled data. The performance of these semi-supervised methods significantly degrades as the size of labeled data decreases. We introduce Mutual-information-based Unsupervised & Semi-supervised Concurrent LEarning (MUSCLE), a hybrid learning approach that uses mutual information to combine both unsupervised and semi-supervised learning. MUSCLE can be used as a stand-alone training scheme for neural networks, and can also be incorporated into other learning approaches. We show that the proposed hybrid model outperforms state of the art on several standard benchmarks, including CIFAR-10, CIFAR-100, and Mini-Imagenet. Furthermore, the performance gain consistently increases with the reduction in the amount of labeled data, as well as in the presence of bias. We also show that MUSCLE has the potential to boost the classification performance when used in the fine-tuning phase for a model pre-trained only on unlabeled data.

In this paper, we study the problem of efficiently assessing building damage after natural disasters like hurricanes, floods or fires, through aerial video analysis. We make two main contributions. The first contribution is a new dataset, consisting of user-generated aerial videos from social media with annotations of instance-level building damage masks. This provides the first benchmark for quantitative evaluation of models to assess building damage using aerial videos. The second contribution is a new model, namely MSNet, which contains novel region proposal network designs and an unsupervised score refinement network for confidence score calibration in both bounding box and mask branches. We show that our model achieves state-of-the-art results compared to previous methods in our dataset.

Lightweight super resolution networks have extremely importance for real-world applications. In recent years several SR deep learning approaches with outstanding achievement have been introduced by sacrificing memory and computational cost. To overcome this problem, a novel lightweight super resolution network is proposed, which improves the SOTA performance in lightweight SR and performs roughly similar to computationally expensive networks. Multi-Path Residual Network designs with a set of Residual concatenation Blocks stacked with Adaptive Residual Blocks: (i) to adaptively extract informative features and learn more expressive spatial context information; (ii) to better leverage multi-level representations before up-sampling stage; and (iii) to allow an efficient information and gradient flow within the network. The proposed architecture also contains a new attention mechanism, Two-Fold Attention Module, to maximize the representation ability of the model. Extensive experiments show the superiority of our model against other SOTA SR approaches.

We present the Multiview Extended Video with Activities (MEVA) dataset, a new and very-large-scale dataset for human activity recognition. Existing security datasets either focus on activity counts by aggregating public video disseminated due to its content, which typically excludes same-scene background video, or they achieve persistence by observing public areas and thus cannot control for activity content. Our dataset is over 9300 hours of untrimmed, continuous video, scripted to include diverse, simultaneous activities, along with spontaneous background activity. We have annotated 144 hours for 37 activity types, marking bounding boxes of actors and props. Our collection observed approximately 100 actors performing scripted scenarios and spontaneous background activity over a three-week period at access-controled venue, collecting in multiple modalities with overlapping and non-overlapping indoor and outdoor viewpoints. The resulting data includes video from 38 RGB and thermal IR cameras, 42 hours of UAV footage, as well as GPS locations for the actors. 122 hours of annotation are sequestered in support of the NIST Activity in Extended Video (ActEV) challenge; the other 22 hours of annotation and the corresponding video are available on our website, along with an additional 306 hours of ground camera data, 4.6 hours of UAV data, and 9.6 hours of GPS logs. Additional derived data includes camera models geo-registering the outdoor cameras and a dense 3D point cloud model of the outdoor scene. The data was collected with IRB oversight and approval and released under a CC-BY-4.0 license.