SOLQ: Segmenting Objects by Learning Queries

Authors: Bin Dong, Fangao Zeng, Tiancai Wang, Xiangyu Zhang, Yichen Wei

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Industry Bin Dong Fangao Zeng Tiancai Wang Xiangyu Zhang Yichen Wei MEGVII Technology {dongbin,zengfangao,wangtiancai,zhangxiangyu,weiyichen}@megvii.com
Pseudocode No The paper describes methods and processes but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github. com/megvii-research/SOLQ.
Open Datasets Yes We validate our method on COCO benchmark [19]. COCO contains 115k images for training, 5k for validation and 20k for testing, involving 80 object categories with instance-level segmentation annotations.
Dataset Splits Yes COCO contains 115k images for training, 5k for validation and 20k for testing, involving 80 object categories with instance-level segmentation annotations.
Hardware Specification Yes All experiments are conducted over 8 Tesla V100 GPUs with batch size 32 except the comparison in Sec. 4.4.
Software Dependencies No The paper mentions software components like ResNet, Deformable DETR, and Adam optimizer but does not provide specific version numbers for these or other key software dependencies.
Experiment Setup Yes For the mask branch, nk is set to 256, hidden dim of MLP is 1024 and λvec is 3.0. Following DETR, λcls = 2, λL1 = 5, λgiou = 2. We train our model with Adam optimizer with weight decay of 1.0 10 4. Models are trained for 50 epochs with the initial learning rate 2.0 10 4 and decayed at 40th epoch by a factor 0.1. Multi-scale training is adopted, where the shorter side is randomly chosen within [408, 800] and the longer side is less or equal to 1333. All experiments are conducted over 8 Tesla V100 GPUs with batch size 32 except the comparison in Sec. 4.4.