Test_Caffe_nets.FasterRCNN_zf fails with ENABLE_FAST_MATH option #24017
Description
System Information
OpenCV: current 4.x (4.8.0-dev)
Platform: Ubuntu 18.04, GCC 7.5
Hardware: Core i5 2500k (no avx2 and no avx512)
Detailed description
Net inference accuracy issue is raised.
Steps to reproduce
cmake -DENABLE_FAST_MATH=1 ../opencv
make -j4
./bin/opencv_test_dnn --gtest_filter="Test_Caffe_nets.FasterRCNN_zf/*"
Log:
[ RUN ] Test_Caffe_nets.FasterRCNN_zf/2, where GetParam() = OCV/CPU
Unmatched prediction: class 7 score 0.986261 box [249.607 x 105.463 from (468.582, 78.9142)]
Highest IoU: 0
/home/alexander/Projects/OpenCV/opencv-master/modules/dnn/test/test_common.impl.hpp:145: Failure
Value of: matched
Actual: false
Expected: true
model name: ZF_faster_rcnn_final.caffemodel
Unmatched reference: class 7 score 0.988779 box [248.791 x 111.586 from (469.849, 75.1756)] IoU diff: 1
/home/alexander/Projects/OpenCV/opencv-master/modules/dnn/test/test_common.impl.hpp:157: Failure
Expected: (refScores[i]) <= (confThreshold), actual: 0.988779 vs 0.8
model name: ZF_faster_rcnn_final.caffemodel
[ FAILED ] Test_Caffe_nets.FasterRCNN_zf/2, where GetParam() = OCV/CPU (2352 ms)
Issue submission checklist
- I report the issue, it's not a question
- I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
- I updated to the latest OpenCV version and the issue is still there
- There is reproducer code and related data files (videos, images, onnx, etc)