Fix inconsistent NMS implementation between CPU and CUDA

[fmassa]

There was an inconsistency in how nms_threshold was handled in NMS between CPU and GPU. More precisely, the difference arises with proposals that satisfy overlap == nms_threshold.

We used ovr >= iou_threshold for CPU

vision/torchvision/csrc/cpu/nms_cpu.cpp

Line 64 in cd17484

if (ovr >= iou_threshold)

and ovr > iou_threshold for CUDA

vision/torchvision/csrc/cuda/nms_cuda.cu

Line 63 in cd17484

if (devIoU<T>(cur_box, block_boxes + i * 4) > iou_threshold) {

This was caught during test refactorings in #1551, which exposed this inconsistency.

cc @rbgirshick @ppwwyyxx

[codecov-io]

Codecov Report

Merging #1556 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1556   +/-   ##
=======================================
  Coverage   65.48%   65.48%           
=======================================
  Files          90       90           
  Lines        7080     7080           
  Branches     1077     1077           
=======================================
  Hits         4636     4636           
  Misses       2135     2135           
  Partials      309      309

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8909ff4...335aa76. Read the comment docs.