Different results for bytetrack graph (custom graph) on Intel CPU and AMD CPU when using OVMS inference

[Vishwa2684]

Hello everyone,

I've implemented the bytetrack algorithm in my fork of openvinotoolkit/mediapipe and these are the results on the 2 of the following devices for the same config of my bytetrack calculator

bytetrack graph config:

• OpenVINOYoloxTensorsToDetections config
  • conf_thresh:0.35
• NonMaxSuppressionCalculator config
  • min_suppression_threshold:0.45
  • max_num_detections: 100
  • overlap_type: INTERSECTION_OVER_UNION
  • return_empty_detections: true
• ByteTrackCalculator config
  • track_high_threshold:0.5
  • track_low_threshold:0.35
  • new_track_threshold:0.25
  • matching_threshold: 0.75
  • track_buffer: 30
  • fuse_score: false

=== OUTPUTS FOR SAME CONFIG ON 2 different devices ===
Device 1:

• CPU: 13th gen Intel(R) Core(TM) i9-13900HX
• GPU 0: Intel(R) UHD Graphics
• GPU 1: NVIDIA GeForce RTX 4060 Laptop GPU

Output:

out_bt_ovms.1.mp4

Device 2:

• CPU: AMD Ryzen 9 7940HS w/ Radeon 780M Graphics
• NPU: NPU Compute Accelerator device
• GPU 0: NVIDIA GeForce RTX 4070 Laptop GPU
• GPU 1: AMD Radeon(TM) Graphics

Output:

out_bt_ovms_dog.mp4

How to reproduce the following issue?

• Download the video from the following link
• Then clone the following repo and checkout to the following f807023 commit in custom_bytetrack_graph and build the docker image and run container.
• Then follow the README to install yolox nano weights and setup the bytetrack demo
• Then build the bytetrack demo according to README and perform inference on the following video in 2 different devices.

[Vishwa2684]

@atobiszei @dtrawins