issamlaradji / affinity_lcfcn Goto Github PK

Hi Issam,

Thank you for posting the great work! I am trying to reproduce the LCFCN results (i.e. MAE, GAME and mIoU) mentioned in your Affinity-LCFCN paper (table 1 and 2). Currently, I am training with the original image resolution 1920*1080 (from DeepFish), which turned out to be very slow (around 18 minutes for one epoch). As mentioned in the A-LCFCN paper, did you resize all the images to (256, 455) during training?

Besides, could you kindly provide your GPU setup and how long was the training for 1000 epochs?

Another question is that you have two sections for miou evaluation: FishLoc and FishSeg. Are you training on FishLoc.train and FishSeg.train respectively, and validating on FishLoc.val and FishSeg.val respectively, and then testing both on FishSeg.test? When validating, how do you pick the best model? Is it based on MAE, GAME, or anything else?

You did mention you used "early stopping with patience of 10 epochs", does it mean you usually stop training before reaching 1000 epochs?

Thank you so much!

Hi,

I tried to reproduce the miou result of LCFCN on the FishSeg test set. No matter how I tried, the best miou for the foreground I have got is around 62%, which is short from the paper results 68.4%. I am curious about the hyperparameter settings you used for this experiment, could you please kindly provide that?

Currently I am using resnet38 as the backbone (pretrained on imagenet, I downloaded from this repo), and I specified the backbone by assigning the "shared" argument True in the exp_dict, even though under the "model" fcn8_vgg16 is used. I use adam and beta (0.9, 0.999). I tried learning rate in {le-4, 1e-5, 1e-6}, and I also tried using a learning rate scheduler. For loss type, I used the lcfcn_loss. But I found there are lots of loss types provided too. I found lcfcn_const_loss can boost the performance too. I use RTX3090 to train these models. However, I still cannot reach the LCFCN performance of 68.4% for the foreground when training only using the point label of the FishSeg data.

Could you please provide more details about your experiment setting? Thank you so much!

Hi there, thanks for this interesting work. However, the fish dataset link can't be accessed here in China. Is it the same as https://cloudstor.aarnet.edu.au/plus/s/NfjObIhtUYO6332 ? I suppose not, since you use different annotations. Can you provide another link that can be easily accessed? Thanks a lot!