WACV 2023 Papers

We propose a novel approach for monocular 3D object detection by leveraging local perspective effects of each object. While the global perspective effect shown as size and position variations has been exploited for monocular 3D detection extensively, the local perspectives has long been overlooked. We propose a new regression target named keyedge-ratios as the parameterization of the local shape distortion to account for the local perspective, and derive the object depth and yaw angle from it. Theoretically, this approach does not rely on the absolute size or position of the objects in the image, therefore independent of the camera intrinsic parameters. This approach provides a new perspective for monocular 3D reasoning and can be plugged in flexibly to existing monocular 3D object detection frameworks. We demonstrate effectiveness and superior performance over strong baseline methods in multiple datasets.

Depth-aware video panoptic segmentation tackles the inverse projection problem of restoring panoptic 3D point clouds from video sequences, where the 3D points are augmented with semantic classes and temporally consistent instance identifiers. We propose a novel solution with a multi-task network that performs monocular depth estimation and video panoptic segmentation. Since acquiring ground truth labels for both depth and image segmentation has a relatively large cost, we leverage the power of unlabeled video sequences with self-supervised monocular depth estimation and semi-supervised learning from pseudo-labels for video panoptic segmentation. To further improve the depth prediction, we introduce panoptic-guided depth losses and a novel panoptic masking scheme for moving objects to avoid corrupting the training signal. Extensive experiments on the Cityscapes-DVPS and SemKITTI-DVPS datasets demonstrate that our model with the proposed improvements achieves competitive results and fast inference speed.

Style may refer to different concepts (e.g. painting style, hairstyle, texture, color, filter, etc.) depending on how the feature space is formed. In this work, we propose a novel idea of interpreting the lighting in the single- and multi-illuminant scenes as the concept of style. To verify this idea, we introduce an enhanced auto white-balance (AWB) method that models the lighting in single- and mixed-illuminant scenes as the style factor. Our AWB method does not require any illumination estimation step, yet contains a network learning to generate the weighting maps of the images with different WB settings. Proposed network utilizes the style information, extracted from the scene by a multi-head style extraction module. AWB correction is completed after blending these weighting maps and the scene. Experiments on single- and mixed-illuminant datasets demonstrate that our proposed method achieves promising correction results when compared to the recent works. This shows that the lighting in the scenes with multiple illuminations can be modeled by the concept of style. Source code and trained models are available on https://github.com/birdortyedi/lighting-as-style-awb-correction.

Scene text erasing aims to wipe text regions in scene images with reasonable background. Most previous approaches employ scene text detectors to assist localization of the text regions. However, detected text boxes contain both text strokes and background clutters, and directly inpainting on the whole boxes may remain text artifacts and make regions unnatural. In this paper, we present an end-to-end network that focuses on modeling text stroke masks that provide more accurate locations to compute erased images. The network consists of two stages, i.e., a basic network with stroke generation and a refinement network with stroke awareness. The basic network predicts the text stroke masks and initial erasing results simultaneously. The refinement network receives the masks as supervision to generate natural erased results. Experiments on both synthetic and real-world scene images demonstrate the effectiveness of our framework in producing high quality erasing results.

In multi-modal action recognition, it is important to consider not only the complementary nature of different modalities but also global action content. In this paper, we propose a novel network, named Modality Mixer (M-Mixer) network, to leverage complementary information across modalities and temporal context of an action for multi-modal action recognition. We also introduce a simple yet effective recurrent unit, called Multi-modal Contextualization Unit (MCU), which is a core component of M-Mixer. Our MCU temporally encodes a sequence of one modality (e.g., RGB) with action content features of other modalities (e.g., depth, IR). This process encourages M-Mixer to exploit global action content and also to supplement complementary information of other modalities. As a result, our proposed method outperforms state-of-the-art methods on NTU RGB+D 60, NTU RGB+D 120, and NWUCLA datasets. Moreover, we demonstrate the effectiveness of M-Mixer by conducting comprehensive ablation studies.

This paper addresses the problem of mobile robot manipulation using object detection. Our approach uses detection and control as complimentary functions that learn from real-world interactions. We develop an end-to-end manipulation method based solely on detection and introduce Task-focused Few-shot Object Detection (TFOD) to learn new objects and settings. Our robot collects its own training data and automatically determines when to retrain detection to improve performance across various subtasks (e.g., grasping). Notably, detection training is low-cost, and our robot learns to manipulate new objects using as few as four clicks of annotation. In physical experiments, our robot learns visual control from a single click of annotation and a novel update formulation, manipulates new objects in clutter and other mobile settings, and achieves state-of-the-art results on an existing visual servo control and depth estimation benchmark. Finally, we develop a TFOD Benchmark to support future object detection research for robotics: https://github.com/griffbr/TFOD.

Many real-world scenarios in which DNN-based recognition systems are deployed have inherently fine-grained attributes (e.g., bird-species recognition, medical image classification). In addition to achieving reliable accuracy, a critical subtask for these models is to detect Out-of-distribution (OOD) inputs. Given the nature of the deployment environment, one may expect such OOD inputs to also be fine-grained w.r.t. the known classes (e.g., a novel bird species), which are thus extremely difficult to identify. Unfortunately, OOD detection in fine-grained scenarios remains largely underexplored. In this work, we aim to fill this gap by first carefully constructing four large-scale fine-grained test environments, in which existing methods are shown to have difficulties. Particularly, we find that even explicitly incorporating a diverse set of auxiliary outlier data during training does not provide sufficient coverage over the broad region where fine-grained OOD samples locate. We then propose Mixture Outlier Exposure (MixOE), which mixes ID data and training outliers to expand the coverage of different OOD granularities, and trains the model such that the prediction confidence linearly decays as the input transitions from ID to OOD. Extensive experiments and analyses demonstrate the effectiveness of MixOE for building up OOD detector in fine-grained environments. The code is available at https://github.com/zjysteven/MixOE.

Visual Place Recognition (VPR) is a crucial part of mobile robotics and autonomous driving as well as other computer vision tasks. It refers to the process of identifying a place depicted in a query image using only computer vision. At large scale, repetitive structures, weather, and illumination changes pose a real challenge, as appearances can drastically change over time. Along with tackling these challenges, an efficient VPR technique must also be practical in real-world scenarios where latency matters. To address this, we introduce MixVPR, a new holistic feature aggregation technique that takes feature maps from pre-trained backbones as a set of global features. Then, it incorporates a global relationship between elements in each feature map in a cascade of feature mixing, eliminating the need for local or pyramidal aggregation as done in NetVLAD or TransVPR. We demonstrate the effectiveness of our technique through extensive experiments on multiple large-scale benchmarks. Our method outperforms all existing techniques by a large margin while having less than half the number of parameters compared to CosPlace and NetVLAD. We achieve a new all-time high recall@1 score of 94.6% on Pitts250k-test, 88.0% on MapillarySLS, and more importantly, 58.4% on Nordland. Finally, our method outperforms two-stage retrieval techniques such as Patch-NetVLAD, TransVPR and SuperGLUE, all while being orders of magnitude faster.

One of the critical steps in biometrics pipeline is detection of presentation attacks, a physical adversary. Several presentation (adversary) attack detection (PAD) algorithms, including iris PAD, are proposed and have shown superlative performance. However, a recent study, on a small-scale database, has highlighted that iris PAD may have gender biases. In this research, we present a rigorous study on gender bias in iris presentation attack detection algorithms using a large-scale and gender-balanced database. The paper provides several interesting observations which can help in building future presentation attack detection algorithms with aim of fair treatment of each demography. In addition, we also present a robust iris presentation attack detection algorithm by combining gender-covariate biased classifiers. The proposed robust classifier not only reduces the difference in accuracy between different genders but also improves the overall performance of the PAD system.

We introduce Meta-OLE, a new geometry-regularized method for fast adaptation to novel tasks in few-shot image classification. The proposed method learns to adapt for each few-shot classification task a feature space with simultaneous inter-class orthogonality and intra-class low-rankness. Specifically, a deep feature extractor is trained by explicitly imposing orthogonal low-rank subspace structures among features corresponding to different classes within a given task. To adapt to novel tasks with unseen categories, we further meta-learn a light-weight transformation to enhance the inter-class margins. As an additional benefit, this light-weight transformation lets us exploit the query data for label propagation from labeled to unlabeled data without any auxiliary network components. The explicitly geometry-regularized feature subspaces allow the classifiers on novel tasks to be inferred in a closed form, with an adaptive subspace truncation that selectively discards non-discriminative dimensions. We perform experiments on standard few-shot image classification tasks, and observe performance superior to state-of-the-art meta-learning methods.

In this paper, we propose an instance-wise meta-learning algorithm for optical flow domain adaptation. Typical optical flow algorithms with deep learning suffer from weak cross-domain performance since their trainings largely rely on synthetic datasets in specific domains. This prevents optical flow performance on different scenes from carrying similar performance in practice. Meanwhile, test-time domain adaptation approaches for optical flow estimation are yet to be studied. Our proposed method, with some training data, learns to adapt more sensitively to incoming inputs in the target domain. During the inference process, our method readily exploits the information only accessible in the test-time. Since our algorithm adapts to each input image, we incorporate traditional unsupervised losses for optical flow estimation. Moreover, with the observation that optical flows in a single domain typically contain many similar motions, we show that our method demonstrates high performance with only a small number of training data. This allows to save labeling efforts. Through the experiments on KITTI and MPI-Sintel datasets, our algorithm significantly outperforms the results without adaptation and shows consistently better performance in comparison to typical fine-tuning with the same amount of data. Also qualitatively our proposed method demonstrates more accurate results for the images with high errors in the original networks.

We consider a new problem of future depth prediction in video. Given a sequence of observed frames, the goal is to predict the depth map of a future frame that has not been observed yet. Depth estimation plays a vital role for scene understanding and decision-making in intelligent systems. Predicting future depth maps can be valuable for autonomous vehicles to anticipate the behaviors of their surrounding objects. Our proposed model for this problem has a two-branch architecture. One branch is for the primary task of future depth estimation. The other branch is for an auxiliary task of image reconstruction. The auxiliary branch can act as a regularization. Inspired by some recent work on test-time adaption, we use the auxiliary task during testing to adapt the model to a specific test video. We also propose a novel meta-auxiliary learning that learn the model specifically for the purpose of effective test-time adaptation. Experimental results demonstrate that our proposed approach significantly outperforms other alternative methods.

Recent advances in synthesizing realistic faces have shown that synthetic training data can replace real data for various face-related computer vision tasks. A question arises: how important is realism? Is the pursuit of photorealism excessive? In this work, we show otherwise. We boost the realism of our synthetic faces by introducing dynamic skin wrinkles in response to facial expressions, and observe significant performance improvements in downstream computer vision tasks. Previous approaches for producing such wrinkles either required prohibitive artist effort to scale across identities and expressions, or were not capable of reconstructing high-frequency skin details with sufficient fidelity. Our key contribution is an approach that produces realistic wrinkles across a large and diverse population of digital humans. Concretely, we formalize the concept of mesh-tension and use it to aggregate possible wrinkles from high-quality expression scans into albedo and displacement texture maps. At synthesis, we use these maps to produce wrinkles even for expressions not represented in the source scans. Additionally, to provide a more nuanced indicator of model performance under deformations resulting from compressed expressions, we introduce the 300W-winks evaluation subset and the Pexels dataset of closed eyes and winks.

Transformers have shown great promise in medical image segmentation due to their ability to capture long-range dependencies through self-attention. However, they lack the ability to learn the local (contextual) relations among pixels. Previous works try to overcome this problem by embedding convolutional layers either in the encoder or decoder modules of transformers thus ending up sometimes with inconsistent features. To address this issue, we propose a novel attention-based decoder, namely CASCaded Attention DEcoder (CASCADE), which leverages the multiscale features of hierarchical vision transformers. CASCADE consists of i) an attention gate which fuses features with skip connections and ii) a convolutional attention module that enhances the long-range and local context by suppressing background information. We use a multi-stage feature and loss aggregation framework due to their faster convergence and better performance. Our experiments demonstrate that transformers with CASCADE significantly outperform state-of-the-art CNN- and transformer-based approaches, obtaining up to 5.07% and 6.16% improvements in DICE and mIoU scores, respectively. CASCADE opens new ways of designing better attention-based decoders.

A match cut is a transition between a pair of shots that uses similar framing, composition, or action to fluidly bring the viewer from one scene to the next. Match cuts are frequently used in film, television, and advertising. However, finding shots that work together is a highly manual and time-consuming process that can take days. We propose a modular and flexible system to efficiently find high-quality match cut candidates starting from millions of shot pairs. We annotate and release a dataset of approximately 20,000 labeled pairs that we use to evaluate our system, using both classification and metric learning approaches that leverage a variety of image, video, audio, and audio-visual feature extractors. In addition, we release code and embeddings for reproducing our experiments at github.com/netflix/matchcut.

Recently, masked image modeling (MIM) has gained considerable attention due to its capacity to learn from vast amounts of unlabeled data and has been demonstrated to be effective on a wide variety of vision tasks involving natural images. Meanwhile, the potential of self-supervised learning in modeling 3D medical images is anticipated to be immense due to the high quantities of unlabeled images, and the expense and difficulty of quality labels. However, MIM's applicability to medical images remains uncertain. In this paper, we demonstrate that masked image modeling approaches can also advance 3D medical images analysis in addition to natural images. We study how masked image modeling strategies leverage performance from the viewpoints of 3D medical image segmentation as a representative downstream task: i) when compared to naive contrastive learning, masked image modeling approaches accelerate the convergence of supervised training even faster (1.40x) and ultimately produce a higher dice score; ii) predicting raw voxel values with a high masking ratio and a relatively smaller patch size is non-trivial self-supervised pretext-task for medical images modeling; iii) a lightweight decoder or projection head design for reconstruction is powerful for masked image modeling on 3D medical images which speeds up training and reduce cost; iv) finally, we also investigate the effectiveness of MIM methods under different practical scenarios where different image resolutions and labeled data ratios are applied. Anonymized codes are available at https://anonymous.4open.science/r/MIM-Med3D.

Hand pose analysis is a key step to understanding dexterous hand performances of many high-level skills, such as playing the piano. Currently, most accurate hand tracking systems are using fabric-/marker-based sensing that potentially disturbs users' performance. On the other hand, markerless computer vision-based methods rely on a precise bare-hand dataset for training, which is difficult to obtain. In this paper, we collect a large-scale high precision 3D hand pose dataset with a small workload using a novel marker-removal network (MR-Net). The proposed MR-Net translates the marked-hand images to realistic bare-hand images, and the corresponding 3D postures are captured by a motion capture system thus few manual annotations are required. A baseline estimation network PiaNet is introduced and we report the accuracy of various metrics together with a blind qualitative test to show the practical effect.

Deep Neural Networks(DNN) often fail in surprising ways, and predicting how well a trained DNN will generalize in a new, external operating domain is essential for deploying DNNs in safety critical applications, e.g., perception for self-driving vehicles or medical image analysis. Recently, the task of Network Generalization Prediction (NGP) has been proposed to predict how a DNN will generalize in an external operating domain. Previous NGP approaches have leveraged multiple labeled test sets or labeled metadata. In this study, we propose an embedding map, the first NGP approach that predicts DNN performance based on how unlabeled images from an external operating domain map in the DNN embedding space. We evaluate our proposed Embedding Map and other recently proposed NGP approaches for pedestrian, melanoma, and animal classification tasks. We find that our embedding map has the best average NGP performance, and that our embedding map is effective at modeling complex, non-linear embedding space structures.

This work presents a novel self-supervised pre-training method to learn efficient representations without labels on histopathology medical images utilizing magnification factors. Other state-of-the-art works mainly focus on fully supervised learning approaches that rely heavily on human annotations. However, the scarcity of labeled and unlabeled data is a long-standing challenge in histopathology. Currently, representation learning without labels remains unexplored in the histopathology domain. The proposed method, Magnification Prior Contrastive Similarity (MPCS), enables self-supervised learning of representations without labels on small-scale breast cancer dataset BreakHis by exploiting magnification factor, inductive transfer, and reducing human prior. The proposed method matches fully supervised learning state-of-the-art performance in malignancy classification when only 20% of labels are used in fine-tuning and outperform previous works in fully supervised learning settings for three public breast cancer datasets, including BreakHis. Further, It provides initial support for a hypothesis that reducing human-prior leads to efficient representation learning in self-supervision, which will need further investigation. The implementation of this work is available online on GitHub.