search

ruoqianguo / DetNet_pytorch

An implementation of DetNet: A Backbone network for Object Detection.
MIT License
377 stars 117 forks source link

A problem about feat_stride #4

Closed rxqy closed 6 years ago

rxqy commented 6 years ago

In your implementation, the feat stride is set to [4, 8, 16, 16, 16] in line 187 of trainval.py FPN_FEAT_STRIDES', '[4, 8, 16, 16, 16] it worked, and it's indeed the case in the original paper, because the feature maps are downsampled 4x, 8x, 16x, 16x, 16x (only dilated, no pooling).

However, from my understanding, as we are using dilated conv, the receptive field of the 3rd, 4 and 5 th layers are enlarged, so I think the stride should be set to [4, 8, 16, 32, 64] ie, one point in the 4th layer represents 2x more information than the 3rd layer. (so 32x?)

Maybe I'm not understanding dilated conv in the right way..? Correct me if I'm wrong.

ruoqianguo commented 6 years ago

@rxqy Using dilated conv can enlarge the receptive field, that's true. But in this code, the FPN_FEAT_STRIDES is just used for generate the anchors in the feature map. you can see the 72 line in proposal_layer_fpn.py. FPN_FEAT_STRIDES just indicates the scale(not the receptive field) between original image and featuremap. Hope it can help you.

rxqy commented 6 years ago

@guoruoqian Thx! Problem solved.