YCBM_seg

Performs pixel level segmentation on RGB images to find YCB-Video/YCBM classes Uses Mobilenet for feature extraction Network Structure based on PoseCNN Obtains pixel level (not IoU) mAR of 67% mAP 64% Dice 85%

Training

Running

python(3) scripts/TwoD_Seg.py [path to training directory] [path to validation directory] [path for model checkpoints] [--optional args]

Directory Structure

└───Train/Val
│   │
│   └───Camera Type 1
|   |   └───Session A
│   |   │   |   000001.jpg (RGB)
│   |   │   |   000001.seg.png (Seg ground truth)
│   |   │   ...
|   |   |  
|   |   └───Session B
│   |   │   |   000001.jpg (RGB)
│   |   │   |   000001.seg.png (Seg ground truth)
│   |   │   ...
|   |   └───Session ...
|   └───Camera Type 2
|   |   └───Session C
|   └───Camera Type ...

Running

python3 main.py path_to_model path_to_image_dir will save results in out_segs.

A few reference images and outputs are provided in images and ref_segs for ease of evaluation

Model weights are available at TBD

Examples