[Paper Review] AFDet: Anchor Free One Stage 3D Object Detection

Runzhou Ge, Zhuangzhuang Ding+5 more|arXiv (Cornell University)|Jun 23, 2020

Advanced Neural Network ApplicationsComputer Science38 references96 citations

TL;DR

AFDet presents an anchor-free, NMS-free one-stage 3D detector for point clouds, enabling embedded-system friendly inference with competitive accuracy on KITTI and Waymo validation sets.

ABSTRACT

High-efficiency point cloud 3D object detection operated on embedded systems is important for many robotics applications including autonomous driving. Most previous works try to solve it using anchor-based detection methods which come with two drawbacks: post-processing is relatively complex and computationally expensive; tuning anchor parameters is tricky. We are the first to address these drawbacks with an anchor free and Non-Maximum Suppression free one stage detector called AFDet. The entire AFDet can be processed efficiently on a CNN accelerator or a GPU with the simplified post-processing. Without bells and whistles, our proposed AFDet performs competitively with other one stage anchor-based methods on KITTI validation set and Waymo Open Dataset validation set.

Motivation & Objective

• Motivate efficient 3D object detection on embedded systems for autonomous driving.
• Eliminate anchor-based and NMS-based post-processing drawbacks for point cloud detectors.
• Develop an end-to-end, anchor-free detector that can run on CNN accelerators or GPUs.
• Show competitive accuracy on standard benchmarks (KITTI and Waymo) compared to single-stage baselines.

Proposed method

• Use PointPillars to encode point clouds into BEV pseudo images, producing a 2D feature map.
• Implement a five-head anchor-free detector: keypoint heatmap, local offset, z-axis location, size, and orientation heads.
• Predict object centers in BEV via a keypoint heatmap and refine centers with a local offset regression map.
• Decode final 3D boxes from BEV centers with z, size, and yaw angle predictions using a structured decoding procedure.
• Replace NMS with a max-pooling and AND-based peak detection in the heatmap for fast, hardware-friendly inference.
• Adopt a modified backbone and neck design to maintain feature map size and reduce computational load while preserving accuracy.

Experimental results

Research questions

• RQ1 Can an anchor-free, NMS-free design achieve competitive 3D detection accuracy on standard benchmarks?
• RQ2 How does an embedded-system friendly design (no NMS, no anchors) impact detection speed and resource usage?
• RQ3 What are the effects of heatmap formulation and offset regression regions on localization and orientation accuracy?
• RQ4 How does AFDet perform on KITTI and Waymo compared to single-stage anchor-based detectors?

Key findings

• AFDet achieves competitive 3D AP on KITTI validation for car detection compared to single-stage baselines.
• A variant of AFDet surpasses state-of-the-art single-stage methods on the Waymo validation set.
• The model is lightweight in parameters and can operate with simplified post-processing suitable for CNN accelerators.
• Replacing traditional NMS with max-pooling and an AND operation yields substantial speedups (embedded-system friendly).
• Heatmap formulation using car-shape prediction and a multi-radius offset improves localization accuracy over center-only offsets.
• On Waymo vehicle detection, AFDet with PointPillars-0.16 beats PointPillars by about 2% on LEVEL_1, and with voxel size 0.10 m outperforms some state-of-the-art single-stage methods.