Shape Prior Guided Instance Disparity Estimation for 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Current 3D object detection from stereo images often relies on computing dense disparity maps for entire images, which is computationally expensive.
• Existing methods fail to effectively utilize category-specific information, limiting the accuracy of disparity estimation.
• Scarcity of annotated disparity data poses a significant challenge for training robust models.

Purpose of the Study:

• To propose a novel system, Disp R-CNN, for accurate 3D object detection using stereo images.
• To develop an instance disparity estimation network (iDispNet) that predicts disparity only for relevant object pixels.
• To enable training without LiDAR data by generating disparity pseudo-ground-truth using a statistical shape model.

Main Methods:

• Introduced an instance disparity estimation network (iDispNet) that learns category-specific shape priors for focused disparity prediction.
• Developed a method to generate dense disparity pseudo-ground-truth using a statistical shape model, overcoming annotation scarcity.
• Integrated iDispNet with a 3D detection framework (R-CNN) for end-to-end 3D object detection.

Main Results:

• Disp R-CNN achieves state-of-the-art performance on the KITTI dataset for 3D object detection from stereo images.
• Demonstrated a 20% improvement in average precision across all categories compared to previous stereo-based methods when LiDAR ground-truth was not used during training.
• The proposed pseudo-ground-truth generation method makes the system widely applicable without requiring expensive LiDAR sensors.

Conclusions:

• Disp R-CNN offers a more efficient and accurate approach to 3D object detection from stereo images by focusing on instance-level disparity estimation.
• The method effectively leverages category-specific priors and overcomes data annotation limitations through synthetic data generation.
• The system's ability to train without LiDAR data significantly broadens its practical applicability in various robotic and autonomous driving scenarios.