mmdetection is very convenient for us to train all kind of detector, but in the work, after training , we are more willing to use the trained detector in our special project to test results. So only forward of a detector is developed in libtorch, written in C++ . do not worry, the c++ code is consistent with python code.

• Python 3.7
• Libtorch 1.3.1
• CUDA: 10.0
• mmdetection: v2.7.0
• mmcv-full: 01.2.1
• CMake

1. SSD
2. FCOS
3. RetinaNet
4. Faster R-CNN
5. Mask R-CNN

reference to mmdetection

• the function forward_trace_xxx( ) is for data stream
• using get_trace_one_stage.py or get_trace_two_stage.py to get traced weights

• using your own traced weights or download my traced weights (9rv5, trained in voc data)
• modifiy the file config/xxx.json
• check that the config file path is correct in example/test_detector.cpp
• just running the code