Normals estimation based on dual camera images.

Examples

0000.png	0001.png	0002.png

Results

Epoch: 39, loss=(total:0.7971 l1:0.3615599572658539 1-cos:0.4355396628379822)
MSE=2.0624 RMSE=1.4332 MAE=1.3626 MME=1.5578
TANGLE11.25=0.0259 TANGLE22.5=0.0725 TANGLE30.0=0.1048

Trained on bdataset_stereo for 40 epochs using Adam.

Train Configuration

LEARNING_RATE: 0.0001
BETAS: [.9, .999]
EPS: 0.00000001
WEIGHT_DECAY: 0.0001

MILESTONES: [10, 20, 30]
GAMMA: .2

NUM_EPOCHS: 40

Google Colab notebook:

References

• Du2Net: Learning Depth Estimation from Dual-Cameras and Dual-Pixels -> Dual camera input
• StereoNet: Guided Hierarchical Refinement for Real-Time Edge-Aware Depth Prediction -> Shared weights for feature extraction in dual camera input
• Depth Estimation via Affinity Learned with Convolutional Spatial Propagation Network -> Uses UNet type architecture
• GeoNet: Geometric Neural Network for Joint Depth and Surface Normal Estimation -> Consistency between depth map and normal map