Skip to content

Latest commit

 

History

History
40 lines (32 loc) · 2 KB

README.md

File metadata and controls

40 lines (32 loc) · 2 KB

Detection with pytorch mobile

Probably first open-source mobile detection on pytorch*

*Let me know in issues if it's not correct

Description
Why it's low-accuracy?
TODO
Links

Description

Most of the detection ops not implemented in pytorch mobile. But it's still possible to implement post-processing ops, like nms, and apply them.

Why SSD has low accuracy?

Checked preprocessing and made so that input tensors in python and java was exactly same. But outputs stays different. Probably, something wrong with flow of tensor of traced SSD model on torch mobile.

TODO

  • Take image from gallery
  • Implement transformations
    • Resize
    • Normalization
  • Implement post-processing
    • Softmax
    • Locations to boxes convertation
    • Non-maximum suppression
  • Draw boxes on image
  • Non real-time object detection (Capture image from camera)
  • Real-time object detection
  • Upload models weights to storage and write downloading script
  • Quantize models (now DETR model params weights 158MB)

Links

Pytorch: Pytorch Mobile
Github: PyTorch Android Examples
Github: Pytorch SSD
Github: Android Camera2 API Example