WACV 2024 论文列表

Most 3D instance segmentation methods exploit a bottom-up strategy, typically including resource-exhaustive post-processing. For point grouping, bottom-up methods rely on prior assumptions about the objects in the form of hyperparameters, which are domain-specific and need to be carefully tuned. On the contrary, we address 3D instance segmentation with a TD3D: the pioneering cluster-free, fully-convolutional and entirely data-driven approach trained in an end-to-end manner. This is the first top-down method outperforming bottom-up approaches in 3D domain. With its straightforward pipeline, it performs outstandingly well on the standard benchmarks: ScanNet v2, its extension ScanNet200, and S3DIS. Besides, our method is much faster on inference than the current state-of-the-art grouping-based approaches: our flagship modification is 1.9x faster than the most accurate bottom-up method, while being more accurate, and our faster modification shows state-of-the-art accuracy running at 2.6x speed. Code is available at https://github.com/SamsungLabs/td3d.

Vision Transformers (ViTs) have emerged as powerful backbones in computer vision, outperforming many traditional CNNs. However, their computational overhead, largely attributed to the self-attention mechanism, makes deployment on resource-constrained edge devices challenging. Multiple solutions rely on token pruning or token merging. In this paper, we introduce "Token Fusion" (ToFu), a method that amalgamates the benefits of both token pruning and token merging. Token pruning proves advantageous when the model exhibits sensitivity to input interpolations, while token merging is effective when the model manifests close to linear responses to inputs. We combine this to propose a new scheme called Token Fusion. Moreover, we tackle the limitations of average merging, which doesn't preserve the intrinsic feature norm, resulting in distributional shifts. To mitigate this, we introduce MLERP merging, a variant of the SLERP technique, tailored to merge multiple tokens while maintaining the norm distribution. ToFu is versatile, applicable to ViTs with or without additional training. Our empirical evaluations indicate that ToFu establishes new benchmarks in both classification and image generation tasks concerning computational efficiency and model accuracy.

The detection of vehicles, pedestrians, and obstacles plays an important role in the decision-making process of autonomous vehicles. While existing methods achieve high detection accuracy under good environmental conditions, they often fail in adverse weather conditions due to limited visibility, blurred contours, and low contrast. These "edge-case" scenarios are not well represented in existing datasets and are not handled properly by object detection algorithms. In our work, we propose a novel approach to synthesising photorealistic and highly diverse scenarios that can be used to fine-tune object detection algorithms in adverse weather conditions such as snow, fog, and rain. The approach uses the Midjourney text-to-image model to create accurate synthetic images of desired weather conditions. Our experiments show that training with our dataset significantly improves detection accuracy in harsh weather conditions. Our results are compared to baseline models and models fine-tuned on augmented clear weather images.

Causal and temporal reasoning about video dynamics is a challenging problem. While neuro-symbolic models that combine symbolic reasoning with neural-based perception and prediction have shown promise, they exhibit limitations, especially in answering counterfactual questions. This paper introduces a method to enhance a neuro-symbolic model for counterfactual reasoning, leveraging symbolic reasoning about causal relations among events. We define the notion of a causal graph to represent such relations and use Answer Set Programming (ASP), a declarative logic programming method, to find how to coordinate perception and simulation modules. We validate the effectiveness of our approach on two benchmarks, CLEVRER and CRAFT. Our enhancement achieves state-of-the-art performance on the CLEVRER challenge, significantly outperforming existing models. In the case of the CRAFT benchmark, we leverage a large pre-trained language model, such as GPT-3.5 and GPT-4, as a proxy for a dynamics simulator. Our findings show that this method can further improve its performance on counterfactual questions by providing alternative prompts instructed by symbolic causal reasoning.

The steady digitization of medieval manuscripts is rapidly changing the field of paleography, challenging existing assumptions about handwriting and book production. This development has identified historically important centers for the production of scribal texts, and even individual scribes themselves. For example, scholars of late medieval English literature have identified the copyists of a number of literary manuscripts, and the important role of London government clerks in shaping literary culture. However, traditional paleography has no agreed-upon methodology or fixed criteria for the attribution of handwriting to a particular community, period, or scribe. The approach taken by paleographers is inherently qualitative and subject to personal bias. Even those wielding the mighty "paleographer's eye" cannot claim objectivity. Computer vision offers solutions with spectacular performance on writer identification and retrieval benchmarks, but these have not been widely adopted by the paleography community because they tend not to hold up in practice. In this work, we attempt to bridge the divide with a software package designed not to automate paleography, but to augment the paleographer's eye. We introduce automated handwriting identification tools for which the results can be quickly visually understood and assessed, and used as one feature among many by expert paleographers when attributing previously unknown scribal hands. We also demonstrate a use case for our software by analyzing several items believed to be written by Thomas Hoccleve, a highly productive clerk of the Privy Seal who also happens to be an important fifteenth-century English poet.

Multiple Object Tracking (MOT) is a rapidly developing research field that targets precise and reliable tracking of objects. Unfortunately, most available MOT datasets typically contain short video clips only, disregarding the indispensable requirement for adequately capturing substantial long-term variations in real-world scenarios. Long-term MOT poses unique challenges due to changes in both the objects and the environment, which remain relatively unexplored. To fill the gap, we propose a time-lapse image dataset inspired by the growth monitoring of strawberries, dubbed "The Growing Strawberries" Dataset (GSD). The data was captured hourly by six cameras, covering a span of 16 months in 2021 and 2022. During this time, it encompassed a total of 24 plants in two separate greenhouses. The changes in appearance, weight, and position during the ripening process, along with variations in the illumination during data collection, distinguish the task from previous MOT research. These practical issues resulted in a drastic performance downgrade in the track identification and association tasks of state-of-the-art MOT algorithms. We believe "The Growing Strawberries" will provide a platform for evaluating such long-term MOT tasks and inspire future research. The dataset is available at https://doi.org/10.4121/e3b31ece-cc88-4638-be10-8ccdd4c5f2f7.v1.

Recent works have shown that objects discovery can largely benefit from the inherent motion information in video data. However, these methods lack a proper background processing, resulting in an over-segmentation of the non-object regions into random segments. This is a critical limitation given the unsupervised setting, where object segments and noise are not distinguishable. To address this limitation we propose BMOD, a Background-aware Motion-guided Objects Discovery method. Concretely, we leverage masks of moving objects extracted from optical flow and design a learning mechanism to extend them to the true foreground composed of both moving and static objects. The background, a complementary concept of the learned foreground class, is then isolated in the object discovery process. This enables a joint learning of the objects discovery task and the object/non-object separation. The conducted experiments on synthetic and real-world datasets show that integrating our background handling with various cutting-edge methods brings each time a considerable improvement. Specifically, we improve the objects discovery performance with a large margin, while establishing a strong baseline for object/non-object separation.

The paper presents CoFormer (Convolutional Fourier Transformer), a robust and adaptable transformer architecture designed for a range of scene text tasks. CoFormer integrates convolution and Fourier operations into the transformer architecture. Thus, it leverages convolution properties such as shared weights, local receptive fields, and spatial subsampling, while the Fourier operation emphasizes composite characteristics from the frequency domain. The research further proposes the first pretraining datasets, named Textverse10M-E and Textverse10M-H. Using these datasets, we demonstrate the efficacy of pretraining for scene text understanding. CoFormer achieves state-of-theart results with and without pretraining on two downstream tasks: scene text recognition and scene text style transfer. The paper presents LISTNet (Language Invariant Style Transfer), a novel framework for bi-lingual scene text style transfer. It also introduces three datasets, viz., TST500K for scene text style transfer, CSTR2.5M and Akshara550 for scene text recognition.

Several recent deep learning (DL) based techniques perform considerably well on image-based multilingual text detection. However, their performance relies heavily on the availability and quality of training data. There are numerous types of page-level document images consisting of information in several modalities, languages, fonts, and layouts. This makes text detection a challenging problem in the field of computer vision (CV), especially for low-resource or handwritten languages. Furthermore, there is a scarcity of word-level labeled data for text detection, especially for multilingual settings and Indian scripts that incorporate both printed and handwritten text. Conventionally, Indian script text detection requires training a DL model on plenty of labeled data, but to the best of our knowledge, no relevant datasets are available. Manual annotation of such data requires a lot of time, effort, and expertise. In order to solve this problem, we propose TEXTRON, a Data Programming-based approach, where users can plug various text detection methods into a weak supervision-based learning framework. One can view this approach to multilingual text detection as an ensemble of different CV-based techniques and DL approaches. TEXTRON can leverage the predictions of DL models pre-trained on a significant amount of language data in conjunction with CV-based methods to improve text detection in other languages. We demonstrate that TEXTRON can improve the detection performance for documents written in Indian languages, despite the absence of corresponding labeled data. Further, through extensive experimentation, we show improvement brought about by our approach over the current State-of-the-art (SOTA) models, especially for handwritten Devanagari text. Code and dataset has been made available at https://github.com/IITB-LEAP-OCR/TEXTRON

This paper addresses the challenge of generating Counterfactual Explanations (CEs), involving the identification and modification of the fewest necessary features to alter a classifier's prediction for a given image. Our proposed method, Text-to-Image Models for Counterfactual Explanations (TIME), is a black-box counterfactual technique based on distillation. Unlike previous methods, this approach requires solely the image and its prediction, omitting the need for the classifier's structure, parameters, or gradients. Before generating the counterfactuals, TIME introduces two distinct biases into Stable Diffusion in the form of textual embeddings: the context bias, associated with the image's structure, and the class bias, linked to class-specific features learned by the target classifier. After learning these biases, we find the optimal latent code applying the classifier's predicted class token and regenerate the image using the target embedding as conditioning, producing the counterfactual explanation. Extensive empirical studies validate that TIME can generate explanations of comparable effectiveness even when operating within a black-box setting.

Diffusion models have demonstrated impressive performance in text-guided image generation. Current methods that leverage the knowledge of these models for image editing either fine-tune them using the input image (e.g., Imagic) or incorporate structure information as additional constraints (e.g., ControlNet). However, fine-tuning large-scale diffusion models on a single image can lead to severe overfitting issues and lengthy inference time. Information leakage from pretrained models also make it challenging to preserve image content not related to the text input. Additionally, methods that incorporate structural guidance (e.g., edge maps, semantic maps, keypoints) find retaining attributes like colors and textures difficult. Using the input image as a control could mitigate these issues, but since these models are trained via reconstruction, a model can simply hide information about the original image when encoding it to perfectly reconstruct the image without learning the editing task. To address these challenges, we propose a text-to-image editing model with an Image Information Removal module (IIR) that selectively erases color-related and texture-related information from the original image, allowing us to better preserve the text-irrelevant content and avoid issues arising from information hiding. Our experiments on CUB, Outdoor Scenes, and COCO reports our approach achieves the best editability-fidelity trade-off results. In addition, a user study on COCO shows that our edited images are preferred 35% more often than prior work.

State-of-the-art face recognition (FR) models often experience a significant performance drop when dealing with facial images in surveillance scenarios where images are in low quality and often corrupted with noise. Leveraging facial characteristics, such as freckles, scars, gender, and ethnicity, becomes highly beneficial in improving FR performance in such scenarios. In this paper, we introduce text-guided face recognition (TGFR) to analyze the impact of integrating facial attributes in the form of natural language descriptions. We hypothesize that adding semantic information into the loop can significantly improve the image understanding capability of an FR algorithm compared to other soft biometrics. However, learning a discriminative joint embedding within the multimodal space poses a considerable challenge due to the semantic gap in the unaligned image-text representations, along with the complexities arising from ambiguous and incoherent textual descriptions of the face. To address these challenges, we introduce a face-caption alignment module (FCAM), which incorporates cross-modal contrastive losses across multiple granularities to maximize the mutual information between local and global features of the face-caption pair. Within FCAM, we refine both facial and textual features for learning aligned and discriminative features. We also design a face-caption fusion module (FCFM) that applies fine-grained interactions and coarse-grained associations among cross-modal features. Through extensive experiments conducted on three face-caption datasets, proposed TGFR demonstrates remarkable improvements, particularly on low-quality images, over existing FR models and outperforms other related methods and benchmarks.

Despite recent progress in static image segmentation, video segmentation is still challenging due to the need for an accurate, fast, and temporally consistent model. Conducting per-frame static image segmentation is not acceptable since it is computationally prohibitive and prone to temporal inconsistency. In this paper, we present bidirectional occlusion-guided feature propagation (BOFP) method with the goal of improving temporal consistency of segmentation results without sacrificing segmentation accuracy, while at the same time keeping the operations at a low computation cost. It leverages temporal coherence in the video by feature propagation from keyframes to other frames along the motion paths in both forward and backward directions. We propose an occlusion-based attention network to estimate the distorted areas based on bidirectional optical flows, and utilize them as cues for correcting and fusing the propagated features. Extensive experiments on benchmark datasets demonstrate that the proposed BOFP method achieves superior performance in terms of temporal consistency while maintaining comparable level of segmentation accuracy at a low computation cost, striking a great balance among the three metrics essential to evaluate video segmentation solutions.

The goal of Referring Video Object Segmentation is to extract an object from a video clip based on a given expression. While previous methods have utilized the transformer's multi-modal learning capabilities to aggregate information from different modalities, they have mainly focused on spatial information and paid less attention to temporal information. To enhance the learning of temporal information, we propose TCE-RVOS with a novel frame token fusion (FTF) structure and a novel instance query transformer (IQT). Our technical innovations maximize the potential information gain of videos over single images. Our contributions also include a new classification of two widely used validation datasets for investigation of challenging cases. Our experimental results demonstrate that TCE-RVOS effectively captures temporal information and outperforms the previous state-of-the-art methods by increasing the J&F score by 4.0 and 1.9 points using ResNet-50 and VSwin-Tiny as the backbone on Ref-Youtube-VOS, respectively, and +2.0 mAP on A2D-Sentences dataset by using VSwin-Tiny backbone. The code is available at https://github.com/haliphinx/TCE-RVOS

Digital human motion synthesis is a vibrant research field with applications in movies, AR/VR, and video games. Whereas methods were proposed to generate natural and realistic human motions, most only focus on modeling humans and largely ignore object movements. Generating task-oriented human-object interaction motions in simulation is challenging. For different intents of using the objects, humans conduct various motions, which requires the human first to approach the objects and then make them move consistently with the human instead of staying still. Also, to deploy in downstream applications, the synthesized motions are desired to be flexible in length, providing options to personalize the predicted motions for various purposes. To this end, we propose TOHO: Task-Oriented Human-Object Interactions Generation with Implicit Neural Representations, which generates full human-object interaction motions to conduct specific tasks, given only the task type, the object, and a starting human status. TOHO generates human-object motions in four steps: 1) it first estimates the object's final position given the task intent; 2) it then generates keyframe poses grasping the objects; 3) after that, it infills the keyframes and generates continuous motions; 4) finally, it applies a compact closed-form object motion estimation to generate the object motion. Our method generates continuous motions that are parameterized only by the temporal coordinate, which allows for upsampling of the sequence to arbitrary frames and adjusting the motion speeds by designing the temporal coordinate vector. This work takes a step further toward general human-scene interaction simulation.

We propose an algorithm for taming Normalizing Flow models - changing the probability that the model will produce a specific image or image category. We focus on Normalizing Flows because they can calculate the exact generation probability likelihood for a given image. We demonstrate taming using models that generate human faces, a subdomain with many interesting privacy and bias considerations. Our method can be used in the context of privacy, e.g., removing a specific person from the output of a model, and also in the context of debiasing by forcing a model to output specific image categories according to a given distribution. Taming is achieved with a fast fine-tuning process without retraining from scratch, achieving the goal in a matter of minutes. We evaluate our method qualitatively and quantitatively, showing that the generation quality remains intact, while the desired changes are applied.

In recent years, there has been a growing emphasis on the intersection of audio, vision, and text modalities, driving forward the advancements in multimodal research. However, strong bias that exists in any modality can lead to the model neglecting the others. Consequently, the model's ability to effectively reason across these diverse modalities is compromised, impeding further advancement. In recent years, there has been a growing emphasis on the intersection of audio, vision, and text modalities, driving forward the advancements in multimodal research. However, strong bias that exists in any modality can lead to the model neglecting the others. Consequently, the model's ability to effectively reason across these diverse modalities is compromised, impeding further advancement. In this paper, we meticulously review each question type from the original dataset, selecting those with pronounced answer biases. To counter these biases, we gather complementary videos and questions, ensuring that no answers have outstanding skewed distribution. In particular, for binary questions, we strive to ensure that both answers are almost uniformly spread within each question category. As a result, we construct a new dataset, named MUSIC-AVQA v2.0, which is more challenging and we believe could better foster the progress of AVQA task. Furthermore, we present a novel baseline model that delves deeper into the audio-visual-text interrelation. On MUSIC-AVQA v2.0, this model surpasses all the existing benchmarks, improving accuracy by 2% on MUSIC-AVQA v2.0, setting a new state-of-the-art performance.

Holistic scene understanding includes semantic segmentation, surface normal estimation, object boundary detection, depth estimation, etc. The key aspect of this problem is to learn representation effectively, as each subtask builds upon not only correlated but also distinct attributes. Inspired by visual-prompt tuning, we propose a Task-Specific Prompts Transformer, dubbed TSP-Transformer, for holistic scene understanding. It features a vanilla transformer in the early stage and tasks-specific prompts transformer encoder in the lateral stage, where tasks-specific prompts are augmented. By doing so, the transformer layer learns the generic information from the shared parts and is endowed with task-specific capacity. First, the tasks-specific prompts serve as induced priors for each task effectively. Moreover, the task-specific prompts can be seen as switches to favor task-specific representation learning for different tasks. Extensive experiments on NYUD-v2 and PASCAL-Context show that our method achieves state-of-the-art performance, validating the effectiveness of our method for holistic scene understanding.

Video composition merges the foreground and background of different videos, presenting challenges due to variations in capture conditions (e.g., saturation, brightness, and contrast). Video harmonization is a vital process in achieving a realistic composite by seamlessly adjusting the foreground's appearance to match the background. In this paper, we propose TSA2, a novel method for video harmonization that incorporates temporal segment adaptation and aggregation. TSA2 divides the inharmonious input sequence into temporal segments, each corresponding to a different frame rate, allowing effective utilization of complementary information within each segment. The method includes the Temporal Segment Adaptation module, which learns and remaps the distribution difference between background and foreground regions, and the Temporal Segment Aggregation module, which emphasizes and aggregates cross-segment information through element-wise correlations. Experimental results demonstrate that TSA2 outperforms advanced image and video harmonization methods quantitatively and qualitatively.