Unsupervised image segmentation based on depth and normal maps clustering

Examples

0000.png	0001.png	0002.png

Results

Epoch: 20, loss=(total:1.8018 c:0.3810 s:1.4208) IOU=0.9285
PRECISION=0.9273
RECALL=1.0000
F1=0.9620

Trained on bdataset for 5 epochs and bdataset_scene for 20 epochs. Train Configuration

LEARNING_RATE: 1e-5
BETAS: [0.9, 0.999]
EPS: 1e-8
WEIGHT_DECAY: 1e-4

MILESTONES: [10]
GAMMA: 0.1

NUM_EPOCHS: 20

Google Colab notebook:

References

• Unsupervised Learning of Image Segmentation Based on Differentiable Feature Clustering