Skip to content

Latest commit

 

History

History
41 lines (33 loc) · 2.17 KB

NOTES.md

File metadata and controls

41 lines (33 loc) · 2.17 KB

Useful links

YOLO models: https://learnopencv.com/mastering-all-yolo-models/ https://learnopencv.com/animal-pose-estimation/ ultralytics/ultralytics#4771 (debug of the keypoints output in a tflite pose model)

DeepLabCut: https://www.nature.com/articles/s41592-022-01443-0

Deer: https://www.mdpi.com/2076-2615/14/18/2640 https://www.mdpi.com/2504-446X/8/10/522

Bears (bearid python): https://bearresearch.org/ https://onlinelibrary.wiley.com/doi/10.1002/ece3.6840

AnimalPose10K https://github.com/AlexTheBad/AP-10K

Android applications:

  • Test1 : Working example of drawings overlay to a previewview -> preparation to represent the outcome of the tensorflow model
  • Test2 : hello world
  • Test3 : Successfully normalized the input TensorImage colors from 255 to 1, and now the right hand is tagged with a red dot.
  • Test5 : working example with different activities, buttons, preview of the camera and image acquisition. Moreover, it works correctly with the AppCompat, see settings in the manifest, lib.version.toml and build.gradle.kts
  • Test6 : working clean example with camera preview
  • YOLOv8PoseApp : working example of image processing with a tflite tensorflow model;
    • The processing is slow, less than 10 Hz. Better to see if it's possible to quantize the model , wether reduce the picture size (less than 640x640)
    • The output seems meaningless, try to understand better the data structure and debug if the input is correct.
    • Try to quantize the model to int8
  • TensorflowModelDebug : working example processing a jpg image. The results are exactly the same provided by the script tflite_debug.py in the pretrained folder of this project. Now it's missing the last part : determine where I can obtain the data I want

YOLO pose model trained with deers

  • try to train a new model with small dataset -> 10 images , 1 epoch. Great effort from the macbook air but absolutely zero good results

TODO :

  • investigate over c++ native interface to speed up the elaborations (very complicated)
  • Generate a set of deer images with a model in blender : different orientations
  • Try dell computer for the model training : CUDA compatibility in linux?
  • train model with virtual dataset