rbgirshick / dp-dpm Goto Github PK

AS BEFORE , I have installed the dp-dpm step by step, and trained the person detector , but how can i used the detector to detect the person on an image?

Hi Ross,
as described in the CVPR paper Deformable Part Models are Convolutional Neural Networks,
there are distance transform pooling layer and object geometry layer in the DPM-CNN.

Are these two layers included in the code?
Thank you.

(Reported to me by Alan Li.)

Details on the issue are sparse. I'm not sure if the problem occurs in training, testing, or both. This issue is primarily meant to document the problem; I likely will not devote time to fixing.

the pool5 features are not fine tuned, but used from the imagenet model directly. Do you happen to have the fine-tuned (on PASCAL VOC 2007) feature pyramid extractor?

Thanks!

At first I meet the same problem: AP=0.00001. So I change the 'cnn_binary_file' from 'CaffeNet.v2.caffemodel' to 'ilsvrc_2012_train_iter_310k', and then I get the right output(AP=53.7). But I can't get the information of parts, so I use the function'pascal_train_parts.m' to train. In this way , the result is wrong. who can tell me, how to get the right result with the use of parts information?