FastMaskRCNN
FastMaskRCNN copied to clipboard
CharlesShang
Almost all predictions are persons (the first class)
I tried prediction with the trained model, however it seems the output are strongly biased towards the first non-background class (person):
The above predictions are obtained from a model trained with 50K iterations.
Other class predictions are not impossible (see the 3rd image, there are a few cars), but it just doesn't seem right to have such a big bias towards person.
Could anyone suggest what shall I be looking at if I wanted to debug this issue? Thanks!
I am not sure about how the original maskrcnn create the target rois for training. From the current version of code, it seems that rois from RPN will be match with the largest IOU ground truth object. However, when the ground truth objects stack on each others, rois is still matched only to the largest IOU object.
@tou wang did you wrote and ranned the test script to get this kind of output?
On Tue, Jun 27, 2017, 16:20 souryuu [email protected] wrote:
I am not sure about how the original maskrcnn create the target rois for training. From the current version of code, it seems that rois from RPN will be match with the largest IOU ground truth object. However, when the ground truth objects stack on each others, rois is still matched only to the largest IOU object.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/CharlesShang/FastMaskRCNN/issues/97#issuecomment-311323060, or mute the thread https://github.com/notifications/unsubscribe-auth/AHMgsysANTLi2fGRhYjOBlBZgF0KxUQRks5sIN5cgaJpZM4OGcMS .
@souryuu Sorry I may have missed your point -- will this ground-truth overlap issue lead to a bias towards person?
@realwecan Hello, Culd you just let me know about the code that you have used for testing the data.Even I have trained the model.But not understanding how to generate the test code.Hope you would consider my request
@realwecan For me it seems that large portion of the images contain person in it (both in term of size and frequency). Also after I trained it up to some point, it seems bias toward not only to person but some particular large and frequently appearing objects.
@souryuu This is strange and could be problematic, as we would like to get rid of this bias if it indeed exists. For how many iterations have you trained your model with so far? I would appreciate if you could also let me know how do your masks look like (e.g., if they look nice/okay/problematic). I wanted to check if the current mask prediction codes are also good.
Can you please share the code of evaluation and trained weights?
@realwecan I am about to get 700k tmr (total training time = 5days). I did not focus on the mask part yet. I saw some of them on tensorboard but they did not look good.
@realwecan @souryuu Hello, This is Nishith.I am Btech student studying in Indian Institute of technology Varanasi.Being new to field of deep learning I would like to request you to provide some direction in the field of testing as the testing code is not provided. Awaiting your reply
@zhuwenzhen I can not share the weight because of the restriction in my working environment. @vidhi09 I do not have the code for testing the network. However if you want to look at the output of bounding box regression and classification during training, you can follow some instruction in #96. @realwecan Could you please share the testing code? It would be nice if we can check segmentation of all bounding boxes at once like in your result.
@souryuu Acutally, I obtained the segmentation masks with your training codes (on validation data, without updating network weights). I will work on a separate testing codes and then share with everyone.
@realwecan could you just elaborate on your idea of visualizing the masks.So I have to give a demo next week with the masks.So If you could provide me the idea with the visualizing of the masks that would be help full.
@souryuu In your issue #96 you have mentioned that "Removing the comment between line 284-291 in train.py should save the image shown above in output/maskrcnn/est_imgs/" but train.py has only around 219 lines. Could you just elaborate on your method.being the beginner i am finding it difficult to understand.Hope you would consider my request.
