Using TensorRT to implement and accelerate YOLO v3. Multi-scale and NMS are included. The acceleration ratio reaches 3 compared to the original darknet. Model: /data/model
Image: /data/images
Build the sample: $ make -j
Run the sample $ ./run.sh
- Upsample layer with nearest-neighbour interpolution. (Interp85 Interp97)
-
solution 1(used): launch reorgOutputKernel to fuse 3 output layers into 1 out layer form, but cost copy time, then do parser and NMS.
-
solution 2(to be implement): iterate every output layer to do parser, then collect all bboxes to do NMS, also cost copy time during collection.
-
solution 3(to be implement): create temp GPU memory to maintain a (float**) variable referring to 3 output layers, then do parser and NMS like ONE layer, based on index relation, but FAKE-ONE layer need to launch kernel 3 times.