Open moothes opened 5 years ago
Moreover, I use the FPN as backbone and achieve 63.25 mAP on 07test set.
I tried fine-tuning the provided CornerNet_500000 on VOC with two GPU for one day, it gives around 81 mAP. Training from scratch on 2 GPUs for one week (250k iters) on VOC gives about 70 mAP according to my preliminary experiments.
I think the reason should be pretrain-weights? for ssd,fssd,rfbssd, they used imagenet pretrained vgg as init weights,but for hourglass-model, there are not pretrained weights.coco dataset is big enough,but voc dataset is not as big as coco
。。so the paper did not try voc07, guess that voc is not big enough to train from scratch
How can i train cornetNet in on VOC data?
I train the source code on 07train+12train sets of PASCAL VOC, and test on 07test set. The mAP is 66.13. Training loss is around 1 while test loss between 2-5. The batch size is set to 11 due to the limitation of our GPU. Other settings are the same as in COCO. I miss something? Anyone has better results? can u tell me how to modify the code to test voc dataset on a train model in voc
i train on voc2007 too but i just get a rusult.json but map, how can i get the map about my model
I train the source code on 07train+12train sets of PASCAL VOC, and test on 07test set. The mAP is 66.13. Training loss is around 1 while test loss between 2-5. The batch size is set to 11 due to the limitation of our GPU. Other settings are the same as in COCO. I miss something? Anyone has better results?