guoxiawang / doobnet Goto Github PK

I noticed in that in src/caffe/layers/class_balanced_sigmoid_cross_entropy_attention_loss_layer.cu that the following code:
bottom_diff[i] = scale[i] * (target_value == 1 ? (1 - sigmoid_data[ i ]) : sigmoid_data[ i ]) * tmp;
suggests that the scale[i] is treated as a constant. It therefore appears to be that only the log(p) and log(1-p) terms carry a gradient and not the Beta^{(1-p)^gamma} or Beta^{p^gamma}.

Is this because it otherwise leads to numerical instability?

https://github.com/xjock/deeplebv3plus-hedplus

Hi, Guoxia Wang. I tried to use your DOOBNet/doobscripts/evaluation/EvaluateOcc.m to evaluate occ edge results. But I don't understand the content of edge_maps, which is in EvaluateBoundary.m

resfile = fullfile(resPath, [imglist{ires}, '.mat']);
edge_maps = load(resfile);
edge_maps = edge_maps.edge_ori;
res_img = zeros([size(edge_maps.edge), 2], 'single');
res_img(:,:,1) = edge_maps.edge;
res_img(:,:,2) = edge_maps.ori;

Can you please tell what is included in edge_maps, and besides, whether the edge_maps.edge and edge_maps.ori are the original result of the test, without any processing(such as NMS)?
Thank you very much!

In the python file doobscript/doobnet_mat2hdf5_edge_ori.py, there are two problems

1. line 343 and 346 should be train_pair_320x320.lst rather than train_pari_320x320.lst
2. line 353 uses Python2 format

print('Down!')
# print 'Down!'

Hi guoxia,

Thanks a lot for your evaluation code, it's very useful! Btw, can you kindly offer a link to edge_ori predictions of previous methods? When I try to run OccCompCurvesPlot.m, I can't find predictions of previous methods. I think the offered source will be a good contribution to the community.

Thank you by advance!

Hi ! Guoxia Wang,
I had read the evaluation matlab code, which include the evaluation method of edge and occlusion orientation. And I found that the part of occlusion orientation eval is a little confused. Here is some piece of code in doobscript/evaluation/collect_eval_bdry_occ.m

So from the code above, the AA_edge tranfer to func:collect_eval_bdry_v2 for edge evaluation, and AA_ori for occlusion orientation evaluation. but just as the annotation I add above, the AA_ori consist of [thresh cntR sumR cntP_occ sumP ], so why using the edge result cntR and sumR here?（I mean, use edge result in orientation eval）; This maybe not proper, and if I eval the result in this way, the orientation result will be influenced by edge result (when the edge result is improve, the orientation result will be improve too)

Besides, in orientation predict task, there is no positive and negative (different from edge task), so it's hard to calculate the "recall" and "precision" as edge task. Why not use the precision to eval the result of orientation map?

Really Thanks for your answer !

Hi !
When I read the code of class_balanced_sigmoid_cross_entropy_attention_loss_layer.cu to figure out the implement detail of this loss, I just found that you use a normalization of "FULL" (sum up the loss and then divide by NHW). But in your paper, you just use a normalization of "BATCH_SIZE" (sum up the loss and then divide by batch_size). So could you please tell me whether methods is proper?
Thanks!

I appreciate it.

Hi GUOxia Wang！recently，I tried to train the ori net of DOC，but it needs .h5 format，so I use your "doobnet_mat2hdf5_edge_ori.py" to tramsform .mat to .h5，and then I get the .h5 training dataset which includes the edge map and ori map channels. But when I train the ori net I encourntered the INF and NAN error in loss. I wonder whether the .h5 dataset can directly apply to DOC or maybe I should make a modification.Thanks for your time!

I'am intersted in your work and want to use this method in my customized dataset but I have no idea how to generate the oritentation label according to the edge label. Can you provide some details about it? Thank your, bro!