Learning Object-Language Alignments for Open-Vocabulary Object Detection

Authors: Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, Jianfei Cai

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two benchmark datasets, COCO and LVIS, demonstrate our superior performance over the competing approaches on novel categories, e.g. achieving 32.0% m AP on COCO and 21.7% mask m AP on LVIS.
Researcher Affiliation Collaboration 1 Monash University 2 Byte Dance 3 The University of Hong Kong
Pseudocode No The paper describes the approach using textual explanations and mathematical formulations but does not include a distinct pseudocode or algorithm block.
Open Source Code Yes Code is available at: https://github.com/clin1223/VLDet.
Open Datasets Yes COCO and COCO Caption. Following open-vocabulary COCO setting (OV-COCO) (Zareian et al., 2021), the COCO-2017 dataset is manually divided into 48 base classes and 17 novel classes, which are proposed by the zero-shot object detection (Bansal et al., 2018). ... For images-text pairs data, we use COCO Caption (Chen et al., 2015) training set, which contains 5 human-generated captions for each image.
Dataset Splits Yes We keep 107,761 images with base class annotations as the training set and 4,836 images with base and novel class annotations as the validation set.
Hardware Specification Yes All the expriments are conducted on 8 NVIDIA V100 GPUs.
Software Dependencies No The paper mentions software components like Faster R-CNN, CLIP, and CenterNet2, but does not provide specific version numbers for these or other relevant software dependencies (e.g., programming language versions, specific library versions).
Experiment Setup Yes In each mini-batch, the ratio of base-class detection data and image-text pair data is 1:4. For the warmup, we increase the learning rate from 0 to 0.002 for the first 1000 iterations. The model is trained for 90,000 iterations using SGD optimizer with batch size 8 and the learning rate is scaled down by a factor of 10 at 60,000 and 80,000 iterations.