search

NVIDIA / DIGITS

Deep Learning GPU Training System
https://developer.nvidia.com/digits
BSD 3-Clause "New" or "Revised" License
4.12k stars 1.38k forks source link

sub 1 accuracy? #1363

Open WERAQS opened 7 years ago

WERAQS commented 7 years ago

hi folks, we've download the kitti database and set up training according to the rules, the machine we have has a Titan X pascal and training takes about 2hr55minutes each time, but as you can see graphs are weird. whats the problem? screenshot from 2016-12-26 21-15-37 screenshot from 2016-12-26 21-16-18

also there are overlapping bounding boxes, many of them. screenshot from 2016-12-26 21-22-40 screenshot from 2016-12-26 21-23-04

gheinrich commented 7 years ago

Hi @WERAQS, apart from the small drop around epoch 8, which is not unusual, your learning curves look OK, with precision and recall both above 78% (remember that you need to use the left axis to read precision, recall and mAP). Here, mAP is the product of precision and accuracy, which is not an accurate representation of "mean Average Precision", as stated on https://github.com/NVIDIA/caffe/issues/234.

You might want to tune your clustering parameters to remove the duplicate bounding boxes. Have a look at the clustering layer doc string and in particular check the parameters of cv.groupRectangles (from OpenCV) are fine for your test images.