Results on the PASCAL VOC dataset

[XiaLiPKU]

I see for validation on the PASCAL VOC dataset, you first train with stride 16, and then finetune with stride 8. But as I understood and experimented, original paper of DeeplabV3 didn't finetune with stride 8 for validation, it just finetune for testing. Which means, for validation, the model only trained on the trainaug set. And for testing, the model is then finetuned on the trainval set with stride 8. So, feel sorry to say, the result you report is not fair enough.