search

as641651 / Faster-RCNN-Torch

Torch version of Faster RCNN model for object detection
6 stars 3 forks source link

Test Performance #1

Open vinodkkurmi opened 6 years ago

vinodkkurmi commented 6 years ago

hi Is there any problem in evaluation code?? I am getting around 94 mAP and 96 recall value. Are you doing is with train data ???

as641651 commented 6 years ago

No. I evaluate with PASCAL VOC07 test set

vinodkkurmi commented 6 years ago

@as641651 Hi Thanks for your response. Can you tell at what iteration you got the mAP 68.2% ????

ataldim commented 6 years ago

Hello. I got 68.2% mAP when testing with the pre-trained faster-rcnn VGG16 model, which as I understand from the description was not trained with this implementation. When training with vgg16 for initialiazation, I get a mAP of only 57-58% with the provided config (70000 iterations). Further training didn't increase the result. Do you get a similar mAP value?

as641651 commented 6 years ago

@ataldim Yes I get something similar when initialized with VGG16. In the original py-faster-rcnn, they append the dataset with flipped images, which I have not done here. I ll mostly take this project up in December again and add this functionality.

as641651 commented 6 years ago

@vinodkkurmi The test performance is for the mode trained with the python version of faster rcnn in the original paper. Further training actually did not improve yet, probably because the training set is just half of the original version without flipped images

vinodkkurmi commented 6 years ago

@as641651 Thanks for your response. It means if we train using pre trained VGG16 model (as py-faster rcnn ) using this code, we get only 57-57%map . If we just evaluate py-faster rcnn model it gives the 68.2%map ??? Am I right? If this is true, then your code is far away from the actual result.It means there is some problem in coding part. Can you share what part is missing so we can include that and try to achieve the actual performance?

as641651 commented 6 years ago

@vinodkkurmi In the original py-faster-rcnn, they append the dataset with flipped images, which I have not done here. Doing this will double the training set.

gfsliumin commented 6 years ago

Hi, in ApplyBoxTransform.lua, given a bounding box with center (xa, ya), width wa, and height ha, and given offsets (tx, ty, tw, th), compute the new bounding box (x, y, w, h) as: x = tx wa + xa but you give the gradients (dx, dy, dw, dh) on the output the gradients on the inputs as: dtx = wa dx Is this right? why not dtx=dx/wa?