Browse Proceedings

Prompt-based Continual Learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning. While arriving at a new session, existing prompt-based continual learning methods usually adapt features from pre-...

Text-to-image diffusion models have advanced towards more controllable generation via supporting various additional conditions (e.g., depth map, bounding box) beyond text. However, these models are learned based on the premise of perfect alignment ...

In this paper, we develop an open-set object detector, called Grounding DINO, by marrying Transformer-based detector DINO with grounded pre-training, which can detect arbitrary objects with human inputs such as category names or referring ...

Slow inference speed is one of the most crucial concerns for deploying multi-view 3D detectors to tasks with high real-time requirements like autonomous driving. Although many sparse query-based methods have already attempted to improve the ...

In the current state of 3D object detection research, the severe scarcity of annotated 3D data, substantial disparities across different data modalities, and the absence of a unified architecture, have impeded the progress towards the goal of ...

The success of deep neural networks (DNNs) in real-world applications has benefited from abundant pre-trained models. However, the backdoored pre-trained models can pose a significant trojan threat to the deployment of downstream DNNs. Numerous ...$$$$

Compressed Image Super-resolution (CSR) aims to simultaneously super-resolve the compressed images and tackle the challenging hybrid distortions caused by compression. However, existing works on CSR usually focus on single compression codec, i.e., ...$\mathrm{}\mathrm{}$

This paper presents LLaVA-Plus (Large Language and Vision Assistants that Plug and Learn to Use Skills), a general-purpose multimodal assistant trained using an end-to-end approach that systematically expands the capabilities of large multimodal ...

Despite the success of large-scale pretrained Vision-Language Models (VLMs) especially CLIP in various open-vocabulary tasks, their application to semantic segmentation remains challenging, producing noisy segmentation maps with mis-segmented ...

Training a temporal action segmentation (TAS) model on long and untrimmed videos requires gathering framewise video annotations, which is very costly. We propose a two-stage active learning framework to efficiently learn a TAS model using only a ...

Texturing 3D humans with semantic UV maps remains a challenge due to the difficulty of acquiring reasonably unfolded UV. Despite recent text-to-3D advancements in supervising multi-view renderings using large text-to-image (T2I) models, issues ...$\mathrm{}\mathrm{}\mathrm{}\mathrm{}$

Recently, the Neural Radiance Field (NeRF) advancement has facilitated few-shot Novel View Synthesis (NVS), which is a significant challenge in 3D vision applications. Despite numerous attempts to reduce the dense input requirement in NeRF, it ...

Existing 2D human pose estimation research predominantly concentrates on well-lit scenarios, with limited exploration of poor lighting conditions, which are a prevalent aspect of daily life. Recent studies on low-light pose estimation require the ...

Group activity detection (GAD) is the task of identifying members of each group and classifying the activity of the group at the same time in a video. While GAD has been studied recently, there is still much room for improvement in both dataset ...

We introduce a Depicted image Quality Assessment method (DepictQA), overcoming the constraints of traditional score-based methods. DepictQA allows for detailed, language-based, human-like evaluation of image quality by leveraging Multi-modal Large ...

Correspondences emerge from large-scale vision models trained for generative and discriminative tasks. This has been revealed and benchmarked by computing correspondence maps between pairs of images, using nearest neighbors on the feature grids. ...

Polygonal mesh reconstruction of a raw point cloud is a valuable topic in the field of computer graphics and 3D vision. Especially to 3D architectural models, polygonal mesh provides concise expressions for fundamental geometric structures while ...

In this paper, we introduce Recon3DMind, an innovative task aimed at reconstructing 3D visuals from Functional Magnetic Resonance Imaging (fMRI) signals, marking a significant advancement in the fields of cognitive neuroscience and computer ...

We present a unified, promptable model capable of simultaneously segmenting, recognizing, and captioning anything. Unlike SAM, we aim to build a versatile region representation in the wild via visual prompting. To achieve this, we train a ...

Predicting realistic ground views from satellite imagery in urban scenes is a challenging task due to the significant view gaps between satellite and ground-view images. We propose a novel pipeline to tackle this challenge, by generating ...

Machine unlearning has become a pivotal task to erase the influence of data from a trained model. It adheres to recent data regulation standards and enhances the privacy and security of machine learning applications. In this work, we present a new ...

Existing neural radiance field-based methods can achieve real-time rendering of small scenes on the web platform. However, extending these methods to large-scale scenes still poses significant challenges due to limited resources in computation, ...

Deep learning-based multi-view facial capture methods have shown impressive accuracy while being several orders of magnitude faster than a traditional mesh registration pipeline. However, the existing systems (e.g. TEMPEH) are strictly restricted ...

In this paper, we study multi-target domain adaptation of scene understanding models. While previous methods achieved commendable results through inter-domain consistency losses, they often assumed unrealistic simultaneous access to images from ...

Besides a 3D mesh, Human Mesh Recovery (HMR) methods usually need to estimate a camera for computing 2D reprojection loss. Previous approaches may encounter the following problem: both the mesh and camera are not correct but the combination of ...

Sparsely Annotated Object Detection (SAOD) tackles the issue of incomplete labeling in object detection. Compared with Fully Annotated Object Detection (FAOD), SAOD is more complicated and challenging. Unlabeled objects tend to provide wrong ...$$

Recently, the emergence of the large-scale vision-language model (VLM), such as CLIP, has opened the way towards open-world object perception. Many works have explored the utilization of pre-trained VLM for the challenging open-vocabulary dense ...