End-to-End Supervised Stereo Imaging-Based Method for Depth Estimation

Proposed Architecture

The proposed approach employs two encoders and one decoder. The encoders extract the features of stereo images, and the decoder reconstructs the depth map.

• An attention module named Spatial and Channel Attention Module (SCAM) is incorporated in the bottleneck to combine and emphasize the most meaningful features of encoders.
• Moreover, for better gradient propagation and faster convergence, the decoder module reuses the feature maps of encoders in two different ways, i.e., (1) concatenation and (2) element-wise addition using non-identity mapping.
• Extensive ablation studies are conducted to evaluate the proposed architecture’s effectiveness and strategies. The experiments are conducted on three publicly available datasets: RGB+D scene, Cityscapes, and KITTI dataset

A copy of my Master Thesis is Available Here