search

szagoruyko / imagine-nn

IMAGINE torch neural network routines
Other
109 stars 35 forks source link

Looking for helps for ROIWarping layer #38

Open lim0606 opened 8 years ago

lim0606 commented 8 years ago

Hi, this is Jaehyun Lim

I'm sure this is not a proper post in this issue session. However, I just leave this post for looking for helps.

I am currently working on re-implementing the winning solution of ILSVRC & MSCOCO 2015 competition (http://arxiv.org/abs/1512.04412).

As a part of the work, I made the ROI warping layer described in the paper. (see PR https://github.com/szagoruyko/imagine-nn/pull/37)

Unfortunately, the layer is not complete though. However, I want to share it so that somebody can find the bugs that I couldn't find it until now.......

Note:

  1. The equation (8) in the paper seems to make no sense for me. Thus, I made bilinear warping/mapping as described in Wikipedia.
  2. The backprop (gradient) w.r.t. Image input(or blob) works fine (confirmed with test_jacobian), but the backprop w.r.t. delta_rois (fast-rcnn style rois re-parameterization w.r.t anchors) couldn't make it.

The backprop w.r.t delta_rois are a bit off from the nn.Jacobian.forward. I have tried to found what I might miss, but I couldn't find it until now.

I would appreciate if someone helps me find the bugs.

Best regards,

Jaehyun

szagoruyko commented 8 years ago

good work @lim0606, have you looked at https://github.com/qassemoquab/stnbhwd implementation ? not exactly the same, but could help you I think

lim0606 commented 8 years ago

@szagoruyko

Thank you for your kind response.

It seems really helpful! I will look into it.

Best regards,

Jaehyun Lim