[v1.x] ONNX: add faster_rcnn_fpn models #20190
Conversation
|
Hey @Zha0q1 , Thanks for submitting the PR
CI supported jobs: [website, clang, centos-cpu, unix-cpu, windows-cpu, miscellaneous, sanity, windows-gpu, edge, unix-gpu, centos-gpu] Note: |
| onnx_class_ids, onnx_scores, onnx_boxes = session.run([], {input_name: img_data.asnumpy()}) | ||
| assert_obj_detetion_result(mx_class_ids[0], mx_scores[0], mx_boxes[0], | ||
| onnx_class_ids[0], onnx_scores[0], onnx_boxes[0], | ||
| score_thresh=0.8, score_tol=0.05, box_tol=30) |
There was a problem hiding this comment.
Other object detection models are able to pass the test with box_tol=0.01, but here we loose it to box_tol=30 for faster_rcnn models. Do we know what causes the accuracy regression here?
There was a problem hiding this comment.
I think the onnx model can predict the same objects with very similar scores. It's just with a few boxes one of the four values (xstart, xend, ystart, yend) can differ by quit a bit (refer to the bbox of the boy in red). I am not sure about the cause, but it must be related to the fpn structure.
)
#20178 fixes 2 faster_rcnn_fpn models too, adding them to model tests