Real Use Cases

Here we publish real use case models built on NetsPresso.

Vehicle Detection (ITS: Intelligent transportation system)

• Dataset: Nota's custom ITS dataset

• Target HW: NVIDIA Jetson Nano

• Model: NP-Tradeoff

• Accuracy: 93.1%

• Latency: 51ms (x2.37)

• NP-Performance, NP-Tradeoff, and NP-Latency have different architecture and image resolution.
• TRT Engine created with trtexec, explictBatach, FP16 option.
• Latency measured with trtexec, 'latencyMS' key.
  
  

Fisheye People Detection

• Dataset: CEPDOF, HABBOF

• Target HW: Rpi 4B, Rpi Zero W

• Model: NP-Tradeoff

• mAP: 91.5%

• Latency Rpi 4B: 0.3s (x3.23)

• Latency Rpi Zero W: 27.9s (x2.68)

Optimized for Rpi 4B

• NP-Performance, NP-Tradeoff, and NP-Latency have different architecture and image resolution.
• Latency measured from TensorFlow Lite benchmark tool result with XNNPACK, FP16 option.

Optimized for Rpi Zero W

• NP-Performance, NP-Tradeoff, and NP-Latency have different architecture and image resolution.
• Latency measured from TensorFlow Lite benchmark tool result with XNNPACK, FP16 option.

Text Localization

• Dataset: Nota's custom text localization dataset

• Target HW: NVIDIA Jetson Xavier NX

• Model: NP-Tradeoff

• Accuracy: 61.5%

• Latency: 51ms (x4.53 faster)

• NP-Performance, NP-Tradeoff, and NP-Latency have different architecture and image resolution.
• TRT Engine created with trtexec, explictBatach, FP16 option.
• Latency measured with trtexec, 'latencyMS' key.