Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Understanding ROIPool

See original GitHub issue

❓ Questions and Help

Shouldn’t feature_map[:,:,:7,:7].max(1)[0] be equal to roi_pool(feature_map[None], bbox_tmp) in the following case? thanks in advance!

import torch
from maskrcnn_benchmark.layers import ROIPool
feature_map = torch.rand(1,256, 80, 160).cuda()
bbox_tmp = torch.FloatTensor([[0, 0, 0, 7, 7]]).cuda()
roi_pool = ROIPool((7,7), spatial_scale=1.0)
roi_pool(feature_map[None], bbox_tmp)                                                                                                                                                                                                               
tensor([[[[0.8403, 0.8633, 0.8704, 0.8704, 0.6135, 0.8561, 0.8561],
          [0.8403, 0.8633, 0.8633, 0.8820, 0.8820, 0.8215, 0.3497],
          [0.8174, 0.7352, 0.8100, 0.8820, 0.8820, 0.8215, 0.9108],
          [0.8174, 0.6335, 0.8100, 0.8100, 0.8296, 0.8296, 0.9108],
          [0.9459, 0.9459, 0.6566, 0.6566, 0.8296, 0.8296, 0.3599],
          [0.9459, 0.9459, 0.8508, 0.6646, 0.7663, 0.7663, 0.6380],
          [0.8416, 0.8508, 0.9013, 0.9013, 0.7663, 0.7663, 0.6380]]]],
       device='cuda:0')

feature_map[:,:,:7,:7].max(1)[0]
tensor([[[0.9961, 0.9978, 0.9998, 0.9906, 0.9947, 0.9965, 0.9963],
         [0.9989, 0.9929, 0.9914, 0.9813, 0.9977, 0.9999, 0.9981],
         [0.9976, 0.9995, 0.9979, 0.9955, 0.9948, 0.9991, 0.9985],
         [0.9992, 0.9991, 0.9988, 0.9970, 0.9972, 0.9950, 0.9986],
         [0.9982, 0.9966, 0.9968, 0.9852, 0.9983, 0.9897, 0.9916],
         [0.9890, 0.9920, 0.9989, 0.9997, 0.9959, 0.9977, 0.9932],
         [0.9985, 0.9861, 0.9852, 0.9946, 0.9983, 0.9993, 0.9971]]],
       device='cuda:0')

Issue Analytics

State:
Created 5 years ago
Comments:6 (3 by maintainers)

Top GitHub Comments

1reaction

IssamLaradjicommented, Jan 11, 2019

Ah that makes sense! thanks a lot 😃

0reactions

fmassacommented, Jan 11, 2019

Your bounding box is going outside of the image, which is causing it to add some padding while performing the crop.

This gives a zero tensor:

feature_map = torch.rand(1, 1, 5, 5).cuda()
bbox = torch.tensor([[0, 0, 0, 4, 4]], dtype=torch.float32, device='cuda')
roi_pool = ROIPool((5, 5), spatial_scale=1.0)
print(roi_pool(feature_map, bbox) - feature_map)
tensor([[[[0., 0., 0., 0., 0.],
          [0., 0., 0., 0., 0.],
          [0., 0., 0., 0., 0.],
          [0., 0., 0., 0., 0.],
          [0., 0., 0., 0., 0.]]]], device='cuda:0')

I’m closing the issue, but let me know if you still have questions