DiPEx: Dispersing Prompt Expansion for Class-Agnostic Object Detection

Authors: Jia S Lim, Zhuoxiao Chen, Zhi Chen, Mahsa Baktashmotlagh, Xin Yu, Zi Huang, Yadan Luo

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of Di PEx through extensive class-agnostic OD and OOD-OD experiments on MS-COCO and LVIS, surpassing other prompting methods by up to 20.1% in AR and achieving a 21.3% AP improvement over SAM.
Researcher Affiliation Academia The University of Queensland {jiasyuen.lim, zhuoxiao.chen, m.baktashmotlagh}@uq.edu.au {zhi.chen, xin.yu, helen.huang, y.luo}@uq.edu.au
Pseudocode Yes Algorithm 1 The Proposed Di PEx for Class-Agnostic Object Detection
Open Source Code Yes The code is available at https://github.com/jason-lim26/Di PEx.
Open Datasets Yes Datasets. We conduct our experiments using two detection datasets: 1). MS-COCO [32], a large-scale object detection and instance segmentation dataset, comprising approximately 115K training images and 5K validation images across 80 classes. 2). LVIS [15] includes 2.2 million high-quality instance segmentation masks covering 1,000 class labels, resulting in a long-tailed data distribution. It consists of around 100K training images and 19.8K validation images.
Dataset Splits Yes Datasets. We conduct our experiments using two detection datasets: 1). MS-COCO [32], a large-scale object detection and instance segmentation dataset, comprising approximately 115K training images and 5K validation images across 80 classes. 2). LVIS [15]... It consists of around 100K training images and 19.8K validation images.
Hardware Specification Yes Our code is developed on the Open Grounding-DINO framework [63], and operates on a single NVIDIA RTX A6000 GPU with 48 GB of memory.
Software Dependencies No The paper states: “Our code is developed on the Open Grounding-DINO framework [63],” but does not provide specific version numbers for this framework or any other software dependencies like Python, PyTorch, etc.
Experiment Setup Yes For our experiments, we choose a batch size of 8 for training, and set hyperparameter γ = 0.1, τp = 0.1, τc = 0.1, θ = 15 , K = 9, L = 3, and while adopting all remaining hyperparameters from the Open Grounding-DINO codebase.