[Paper Review] RefinedMPL: Refined Monocular PseudoLiDAR for 3D Object Detection in Autonomous Driving

Jean Marie Uwabeza Vianney, Shubhra Aich+1 more|arXiv (Cornell University)|Nov 21, 2019

Advanced Neural Network Applications56 references42 citations

TL;DR

This paper introduces unsupervised and supervised sparsification of dense monocular PseudoLiDAR to improve 3D object detection, achieving state-of-the-art results on KITTI Car and Pedestrian datasets."

ABSTRACT

In this paper, we strive for solving the ambiguities arisen by the astoundingly high density of raw PseudoLiDAR for monocular 3D object detection for autonomous driving. Without much computational overhead, we propose a supervised and an unsupervised sparsification scheme of PseudoLiDAR prior to 3D detection. Both the strategies assist the standard 3D detector gain better performance over the raw PseudoLiDAR baseline using only ~5% of its points on the KITTI object detection benchmark, thus making our monocular framework and LiDAR-based counterparts computationally equivalent (Figure 1). Moreover, our architecture agnostic refinements provide state-of-the-art results on KITTI3D test set for "Car" and "Pedestrian" categories with 54% relative improvement for "Pedestrian". Finally, exploratory analysis is performed on the discrepancy between monocular and LiDAR-based 3D detection frameworks to guide future endeavours.

Motivation & Objective

• Address the high background density and computational burden of raw PseudoLiDARs in monocular 3D detection.
• Propose unsupervised and supervised preprocessing pipelines to refine PseudoLiDAR before 3D detection.
• Demonstrate performance gains on KITTI 3D object detection benchmark.
• Provide insights into the gap between monocular and LiDAR-based 3D detection.
• Evaluate data-engineering trade-offs with architecture-agnostic refinements.

Proposed method

• Unsupervised sparsification: Points of Interest Selection (PoIS) via LoG extrema on forward difference image; Approximate Foreground Separation (AFgS) using 3D nearest-neighbor queries; Distance-Stratified Sampler (DSS) to preserve foreground density while reducing points.
• Supervised sparsification: Depth-Shared 2D Detector (DSD) that shares encoder features from a monocular depth estimator to produce foreground maps; DSS as above.
• Both pipelines reduce dense PseudoLiDAR by ~95% while maintaining or improving detection accuracy.
• Detector: PointRCNN used as the 3D detector to leverage point-space processing.
• Evaluation uses KITTI3D with AP|R40 metric to compare against recent literature.

Experimental results

Research questions

• RQ1 Can sparsifying dense monocular PseudoLiDAR prior to 3D detection improve detection accuracy and reduce computation?
• RQ2 How do unsupervised and supervised sparsification strategies compare in terms of performance and practicality?
• RQ3 What is the impact of PseudoLiDAR density on 3D detectors, and how does distance-based sampling affect small/distant objects?
• RQ4 To what extent can monocular depth priors be leveraged to enhance 3D object detection without additional heavy models?

Key findings

• Unsupervised sparsification and supervised sparsification both outperform the raw PseudoLiDAR baseline on KITTI3D Car validation.
• Supervised approach yields the best KITTI leaderboard performance for Car and Pedestrian categories.
• Sparse Refinement reduces background point density and resolves many false positives caused by overly dense PseudoLiDAR.
• On KITTI3D test set, RefinedMPL achieves state-of-the-art Car and Pedestrian results with notable gains for Pedestrian (54% relative improvement).
• Depth-shared features (DSD) provide competitive 2D foreground cues without substantial computational overhead.