Auto Seg-Loss: Searching Metric Surrogates for Semantic Segmentation

Authors: Hao Li, Chenxin Tao, Xizhou Zhu, Xiaogang Wang, Gao Huang, Jifeng Dai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on PASCAL VOC and Cityscapes demonstrate that the searched surrogate losses outperform the manually designed loss functions consistently.
Researcher Affiliation Collaboration Hao Li1 , Chenxin Tao2 , Xizhou Zhu3, Xiaogang Wang1,3, Gao Huang2, Jifeng Dai3,4 1The Chinese University of Hong Kong 2Tsinghua University 3Sense Time Research 4Qing Yuan Research Institute, Shanghai Jiao Tong University
Pseudocode Yes Algorithm 1: Auto Seg-Loss Parameter Search
Open Source Code Yes Code shall be released at https://github.com/fundamentalvision/Auto-Seg-Loss.
Open Datasets Yes We evaluate on the PASCAL VOC 2012 (Everingham et al., 2015) and the Cityscapes (Cordts et al., 2016) datasets.
Dataset Splits Yes During the surrogate parameter search, we randomly sample 1500 training images in PASCAL VOC and 500 training images in Cityscapes to form the hold-out set Shold-out, respectively. The remaining training images form the training set Strain in search.
Hardware Specification Yes The search time is counted on 8 NVIDIA Tesla V100 GPUs.
Software Dependencies No The paper mentions deep learning models (e.g., Deeplabv3+, Res Net-50/101, PSPNet, HRNet), optimization algorithms (SGD), and a search algorithm (PPO2), but it does not specify version numbers for any of the underlying software frameworks or libraries (e.g., PyTorch, TensorFlow, scikit-learn versions).
Experiment Setup Yes During the surrogate parameter search, we randomly sample 1500 training images in PASCAL VOC and 500 training images in Cityscapes to form the hold-out set Shold-out, respectively. The remaining training images form the training set Strain in search. µ0 is set to make g(y; θ) = y. The backbone network is Res Net-50. The images are down-sampled to be of 128 128 resolution. SGD lasts only 1000 iterations with a mini-batch size of 32.