search

MoonBlvd / tad-IROS2019

Code of the Unsupervised Traffic Accident Detection paper in Pytorch.
MIT License
164 stars 39 forks source link

Flow Features #7

Closed srikanthmalla closed 5 years ago

srikanthmalla commented 5 years ago

Hi Brian, Could you point me to the script where you did ROIPooling of flow features and saved it for different datasets?

Thank you, Srikanth

MoonBlvd commented 5 years ago

@srikanthmalla Yes, if you go to lib/utils/data_prep_utils.py there is roi_pooling_opencv function, which might be what you are looking for.

srikanthmalla commented 5 years ago

1 Did you try ROI pooling of other feature layers instead of output flow image?

  1. Can bilinear interpolation be called as ROI pooling, I think it should be called ROI Align to avoid confusion? https://everitt257.github.io/blog/2019/02/07/RoI-Explained.html
MoonBlvd commented 5 years ago
  1. Not as I remember. But I tried pool from intermediate layers of semantic segmentation and it didn't help.
  2. I think the bilinear interpolation in ROIAlign is different about what we did. If my understanding is correct, ROIAlign used bilinear to solve the problem that the ROIs doesn't align with the conv layer output size. For example if you have a CH3232 output but your ROI vertices in original image, e.g. 12801280, is (500, 600), which cannot be aligned to any coordinate on the 3232 feature map. But what we did was to use bilinear interpolation during resizing, which is simpler and not related to alignment issue. In fact we simply rounded our scaled ROI coordinates to align with the flow map we are pooling from, simple but effective.