Disp R-CNN

Disp R-CNN predicts per-instance disparities from stereo images to enable accurate 3D object detection by integrating category-specific shape priors into instance disparity estimation.

Key Features:

• Instance Disparity Estimation Network (iDispNet): iDispNet predicts disparities only for pixels associated with object instances rather than computing full-image disparity maps, reducing computation and focusing estimation on objects of interest.
• Category-Specific Shape Prior: The network integrates category-specific shape priors to constrain and improve disparity predictions tailored to object categories.
• Statistical Shape Model for Pseudo-Ground-Truth Generation: A statistical shape model generates dense pseudo-ground-truth disparity data without relying on LiDAR point clouds, enabling training with limited annotated disparities.

Scientific Applications:

• 3D Object Detection from Stereo Images: Provides per-instance disparity estimates for 3D localization and bounding-box inference from stereo imagery.
• Autonomous Driving: Supports detection and localization of vehicles and pedestrians using stereo camera inputs for driving scenarios.
• Robotics and Augmented Reality: Supplies instance-level depth information for scene understanding and interaction in robotic perception and AR systems.
• Benchmarking and Evaluation: Applicable for evaluation on stereo-based 3D detection benchmarks such as the KITTI dataset.

Methodology:

Disp R-CNN uses iDispNet to predict disparities for object-associated pixels, integrates category-specific shape priors to guide those predictions, and employs a statistical shape model to generate dense pseudo-ground-truth disparities without LiDAR.