This is an offical implementation of the paper CAR: Class-aware Regularizations for Semantic Segmentation:
@inproceedings{cCAR,
author = {Ye Huang and Di Kang and Liang Chen and Xuefei Zhe and Wenjing Jia and Linchao Bao and Xiangjian He},
title = {CAR: Class-aware Regularizations for Semantic Segmentation},
booktitle = {ECCV},
year = {2022},
}
July-12-2022 : From July 12-2022, the compressed tfrecord is used for Pascal Context. Please convert Pascal Context again by following docs (Only for prior users).
July-4-2022 : CAR: Class-aware Regularizations for Semantic Segmentation has been accepted by ECCV 2022.
May-7-2022 : There were few documentation errors in previous commits and we fixed them today. Sorry for any inconvenience caused.
- Install TensorFlow (≥2.8) + Pillow + tqdm (docs)
- Download and perpare the datasets (docs)
- Download pretrained backbone weights (docs)
- Train (docs)
- Tell me if you successfully reproduced our result :)
Exactly the same results should be obtained if you are using 8 × NVIDIA V100 (SXM2) with iseg <= 0.04. We verified this on many different machines. Note that, you have to use all GPUs on the machine to avoid a deterministic bug that is still under investigation.
To help verify the exact reproduction process, a training log of ResNet-50 + Self-Attention + CAR is provided in resnet50_sa_car_train_log.md