human/TODO.md

2.8 KiB

To-Do list for Human library

Work-in-Progress



Exploring



Known Issues & Limitations

Face with Attention

FaceMesh-Attention is not supported when using WASM backend due to missing kernel op in TFJS
No issues with default model FaceMesh

Object Detection

NanoDet model is not supported when using WASM backend due to missing kernel op in TFJS
No issues with default model MB3-CenterNet

WebGPU

Experimental support only until support is officially added in Chromium
Enable via chrome://flags/#enable-unsafe-webgpu

Firefox

Running in web workers requires OffscreenCanvas which is still disabled by default in Firefox
Enable via about:config -> gfx.offscreencanvas.enabled
Details

Safari

No support for running in web workers as Safari still does not support OffscreenCanvas
Details



Pending Release Changes

Optimizations:

  • Enabled high-resolution optimizations
    Internal limits are increased from 2k to 4k
  • Enhanced device capabilities detection
    See human.env.[agent, wasm, webgl, webgpu] for details
  • If config.backend is not set, Human will auto-select best backend
    based on device capabilities
  • Enhanced support for webgpu

Features:

  • Add draw label templates
  • Add config.filter.autoBrightness (enabled by default)
    Per-frame video on-the-fly brightness adjustments
    Which significantly increases performance and precision in poorly lit scenes
  • Improved config.filter.equalization (disabled by default)
    Image and video on-demand histogram equalization

Architecture:

  • Reduce build dependencies
    Human is now 30% smaller :)
    As usual, Human has zero runtime dependencies,
    all devDependencies are only to rebuild Human itself
  • Upgrade to TFJS 4.0 with strong typing
    see notes on how to use
  • TypeDef refactoring
  • Add named export for improved bundler support when using non-default imports