We use 28 classes for body-part segmentation along with the background class. You can checkout more details on the classes here.
0: Background
1: Apparel
2: Face_Neck
3: Hair
4: Left_Foot
5: Left_Hand
6: Left_Lower_Arm
7: Left_Lower_Leg
8: Left_Shoe
9: Left_Sock
10: Left_Upper_Arm
11: Left_Upper_Leg
12: Lower_Clothing
13: Right_Foot
14: Right_Hand
15: Right_Lower_Arm
16: Right_Lower_Leg
17: Right_Shoe
18: Right_Sock
19: Right_Upper_Arm
20: Right_Upper_Leg
21: Torso
22: Upper_Clothing
23: Lower_Lip
24: Upper_Lip
25: Lower_Teeth
26: Upper_Teeth
27: Tongue
The body-part segmentation model checkpoints are available at,
| Model | Checkpoint Path |
|---|---|
| Sapiens-0.3B | $SAPIENS_LITE_CHECKPOINT_ROOT/seg/checkpoints/sapiens_0.3b/sapiens_0.3b_goliath_best_goliath_mIoU_7673_epoch_194_$MODE.pt2 |
| Sapiens-0.6B | $SAPIENS_LITE_CHECKPOINT_ROOT/seg/checkpoints/sapiens_0.6b/sapiens_0.6b_goliath_best_goliath_mIoU_7777_epoch_178_$MODE.pt2 |
| Sapiens-1B | $SAPIENS_LITE_CHECKPOINT_ROOT/seg/checkpoints/sapiens_1b/sapiens_1b_goliath_best_goliath_mIoU_7994_epoch_151_$MODE.pt2 |
- Navigate to your script directory:
cd $SAPIENS_LITE_ROOT/scripts/demo/[torchscript,bfloat16]
- For part segmentation (uncomment your model config line):
./seg.sh
Define INPUT for your image directory and OUTPUT for results.
The predictions will be visualized as (.jpg or .png) files, the foreground boolean masks and segmentation probabilities will be stored as .npy files in OUTPUT directory.
These .npy will be used in depth and surface normal visualization.
Adjust BATCH_SIZE, JOBS_PER_GPU, TOTAL_GPUS and VALID_GPU_IDS for multi-GPU configurations.
Note, we skip class label visualization as text on the image in interest of speed.
