Generative Active Learning for Long-tailed Instance Segmentation

Authors: Muzhi Zhu, Chengxiang Fan, Hao Chen, Yang Liu, Weian Mao, Xiaogang Xu, Chunhua Shen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that BSGAL outperforms the baseline approach and effectually improves the performance of longtailed segmentation.
Researcher Affiliation Collaboration 1Zhejiang University, China 2The University of Adelaide, Australia 3The Chinese University of Hong Kong, China 4Ant Group.
Pseudocode Yes Algorithm 1 Pipeline for copy-paste baseline
Open Source Code No The paper does not explicitly state that its own implementation code for BSGAL is open-source or provide a link to a repository for its method.
Open Datasets Yes We choose LVIS (Gupta et al., 2019) as the dataset for our experiments. [...] In this section, we conduct a toy experiment on CIFAR-10 (Krizhevsky et al., 2009)...
Dataset Splits Yes We use the official LVIS dataset splits, with about 100,000 images in the training set and 20,000 images in the validation set.
Hardware Specification Yes And we use 4 Nvidia 4090 GPUs with a batch size of 16 during training. As for Swin-L, [...] And we use 4 Nvidia A100 GPUs with a batch size of 16 for training.
Software Dependencies Yes We first use Stable Diffusion V1.5 (Rombach et al., 2022a) (SD) as the generative model.
Experiment Setup Yes For Res Net-50, the maximum training iteration is set to 90,000 [...] The other unspecified parameters also follow the same settings as X-Paste (Zhao et al., 2023), such as the Adam W (Loshchilov and Hutter, 2017) optimizer with an initial learning rate of 1e 4.