Parameters for Pascal VOC segmentation

[bermanmaxim]

Hello,

I see you shared the training code for semantic segmentation and parameters used for the Cityscapes dataset for tuning from A1.

Could you please also share which settings where used for training voc_rna-a1_cls21_s8_ep-0001.params and voc_rna-a1_cls21_s8_coco_ep-0001.params from A1? Was it also with random cropping (or simple resize like at test-time), and is the number of epochs similar to the one of cityscapes?

Thank you

[huaxinxiao]

Same. I have tried to re-train the VOC model with 120 epochs (fixed learning rate). It seems the model is over-fitted after 90 epochs (77.65% on VOC val dataset). Then, I selected the No.90 epochs to continue training with linear learning rate (60 epochs). However, the performance doesn't improve much even worse after 20 epochs.

[zhui064]

@huaxinxiao I have used the pretrained model voc_rna-a1_cls21_s8_coco_ep-0001.params to test , but in PASCAL VOC validation dataset, I can only got about 76% mean IoU , which was reported 82.86% in parper and github. did u have similar question. and can u please tell me what is your training settings to train the model which got 77.65% on voc validation dataset?

[shipengai]

Hi, all.@bermanmaxim @huaxinxiao @zhui064. How about the result on VOC val dataset of your own training . Is IoU beyond 79%?

[czzerone]

@huaxinxiao hello, can you provide your train command for the VOC model，did you train the model with the SBD dataset?