Performance Benchmark (MS COCO)

• Dataset: MS COCO

• Target HW: NVIDIA Jetson Nano

(x1.79)

(x1.86)

(x1.44)

(x1.07)

* NP-Performance, NP-Tradeoff, and NP-Latency have different architecture and image resolution.

* Training epoch: 300

* All other parameters about training, we followed the same configuration as the default config of 'from scratch training' in the Github.

* TRT Engine created with trtexec, explictBatach, FP16 option.

* Latency measured with trtexec, 'latencyMS' key.