Different detection/counting performance on jetson nano vs desktop GPU with same weights and configuration

[AngelaYZhang]

Current Behavior

I'm seeing different object detections/counting using Jetson nano vs. desktop GPU using the same yolov4-tiny config and yolov4-tiny weights.

Output video: Nano result Nano terminal output Desktop result Desktop terminal output

How to Reproduce

Running with the command python3 app.py --input-uri ./videos/AVG-TownCentre-raw.webm --mot Using the standard YOLOv4-tiny weights and configuration, set the 80 classes for the _label_map, modified cfg/mot.json 'yolo_detector_cfg' as "YOLOv4Tiny" and set the class_ids to [0]

Describe what you want to do

1. What input videos you will provide, if any: Raw input video here

Your Environment

• Desktop
  • Operating System and version: Ubuntu 18.04
  • NVIDIA Driver version:
  • Used the docker image? yes
• NVIDIA Jetson
  • Which Jetson? Jetson Nano
  • Jetpack version: Jetpack 4.6
  • Ran install_jetson.sh? Yes
  • Reinstalled OpenCV from Jetpack? No

Thanks!