Would you mind introducing your own vehicle detector?
Closed this issue · 1 comments
Since the images of CityFlow-ReID are padded with extra pixels, more background information is captured. To decrease the interference of background, we refine the images with tight bounding boxes generated by our own vehicle detector.
Why can your detector generate tight bounding boxes?
I find the explanation the paper:
We use a multi-stage cascade R-CNN [3] to build our
detection framework, which adopts SENet [11] as the back-
bone feature extractor. To increase the global context in-
formation in the extracted features, we add FPN [17] to the
backbone. The RoIAlign [10] is replaced by deformable
RoI pooling [5] to encounter the camera distortion problem.
Multi-scale and data flipping are exploited as our data aug-
mentation for training, and only multi-scale is employed for
testing. As for the post-processing, softNMS [2] is used to
further boost the detection recall performance. The detec-
tion training dataset is comprised of COCO [18], KITTI [7]
and AICity [28].