because there is no GPU, so in the CPU training their own data, the middle encountered a variety of pits, fortunately did not give up, This process is documented in this article.
1, under the CPU configuration faster r-cnn, reference blog: http://blog.csdn.net/wjx2012yt/article/details/52197698#quote
2, in the CPU training data set, need to py-faster-rcnn within the Roi_pooling_layer and Smooth_l1_loss_layer changed to the CPU version,
and recompile. The blogger has modified it to be replaced directly: http://blog.csdn.net/qq_14975217/article/details/51495844
3, the production of their own VOC data set, this part of the reference blog: http://blog.csdn.net/gvfdbdf/article/details/52214008
4, in order to prevent mixing with the previous model, before training to remove the output folder (or change the other name), but also to the Py-faster-rcnn/data/cache files and
File deletion (if any) in the Py-faster-rcnn/data/vocdevkit2007/annotations_cache.
5, before the training of the changes. My dataset is for people in the monitoring video, so there are only backgrounds and people in both categories. Just follow this blog to make changes:
http://blog.csdn.net/sinat_30071459/article/details/51332084
Also modify the solve file settings in py-faster-rcnn/models/pascal_voc/zf/faster_rcnn_alt_opt, including STEPSIZE,BASE_LR,
Otherwise, it is very easy to loss=-nan when training.
6. Training in PY-FASTER-RCNN
0 ZF PASCAL_VOC
7. Testing
Copy the Py-faster-rcnn\output\faster_rcnn_alt_opt\***_trainval ZF Caffemodel from the training to
Py-faster-rcnn\data\faster_rcnn_models, modify the py-faster-rcnn\tools\demo.py (mainly category modification
And the modification of the test picture). Run
Python demo.py--cpu
8. Results
Maybe I have too little data and bad quality (1000 photos), the number of training is not enough, the detection effect is not very good, the default threshold value can not be detected.
When the demo.py in the Conf_thresh down to 0.1 only to detect a piece, and the speed is very slow, about 13 seconds a sheet, the effect is general.
Faster R-CNN Train your data in a CPU configuration