ambiser
/
human
mirror of https://github.com/vladmandic/human
9
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

added model changes notes

master
Vladimir Mandic 2021-05-30 09:26:06 -04:00
parent 317a8fc76c
commit 78e6de4516
1 changed files with 29 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
Models.md

@ -13,14 +13,39 @@ Default models in Human library are:
- **Hand Analysis**: MediaPipe Hands - **Hand Analysis**: MediaPipe Hands
- **Object Detection**: MobileNet-v3 with CenterNet - **Object Detection**: MobileNet-v3 with CenterNet
## Notes <br>
## Changes
All models are modified from original implementation in following manner:
- Input pre-processing: image enhancements, normalization, etc.
- Caching: custom caching operations to bypass specific model runs when no changes are detected
- Output parsing: custom analysis of HeatMaps to regions, output values normalization, etc.
- Output interpolation: custom smoothing operations
- Model modifications:
- Model definition: reformatted for readability, added conversion notes and correct signatures
- Model weights: quantized to 16-bit float for size reduction
Models are not re-trained so any bias included in the original models is present in `Human`
*For any possible bias notes, see specific model cards*
<br>
## Using Alternatives
`Human` includes implementations for several alternative models which are normally not 1:1 replacement,
but can be switched on-the-fly due to standardized output implementation
<br>
**Body detection** can be switched from `PoseNet` to `BlazePose`, `EfficientPose` or `MoveNet` depending on the use case: **Body detection** can be switched from `PoseNet` to `BlazePose`, `EfficientPose` or `MoveNet` depending on the use case:
- `PoseNet`: Works with multiple people in frame, works with only partial people - `PoseNet`: Works with multiple people in frame, works with only partial people
Best described as works-anywhere, but not with great precision Best described as works-anywhere, but not with great precision
- `MoveNet`: Works with single person in frame, works with only partial people - `MoveNet-Lightning`: Works with single person in frame, works with only partial people
Modernized and optimized version of PoseNet with different model architecture Modernized and optimized version of PoseNet with different model architecture
- `MoveNet-Thunder`: Variation of `MoveNet` with higher precision but slower processing
- `EfficientPose`: Works with single person in frame, works with only partial people - `EfficientPose`: Works with single person in frame, works with only partial people
Experimental model that shows future promise but is not ready for wide spread usage due to performance Experimental model that shows future promise but is not ready for wide spread usage due to performance
- `BlazePose`: Works with single person in frame and that person should be fully visibile - `BlazePose`: Works with single person in frame and that person should be fully visibile
@ -33,6 +58,8 @@ Default models in Human library are:
- `Age Detection`: SSR-Net Age IMDB - `Age Detection`: SSR-Net Age IMDB
- `Face Embedding`: BecauseofAI MobileFace Embedding - `Face Embedding`: BecauseofAI MobileFace Embedding
**Object detection** can be switched from `mb3-centernet` to `nanodet`
<br><hr><br> <br><hr><br>
## List of all models included in Human library ## List of all models included in Human library