MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection

Introduction

This is the official implementation of MonoTAKD which utilizes OpenPCDet for the KITTI dataset.

Paper

MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection (arXiv, Sup. included)

News

[2025/2/27]: MonoTAKD has been accepted by CVPR 2025 🔥🔥🔥

• Release code and pre-trained models for the KITTI dataset.
• Visualization utils are provided to visualize detection results in both camera perspective and BEV perspective.
• MonoTAKD DEMO images & videos are included in this release.

Notice: Due to the short schedule, instructions and pre-trained models will be released gradually in the near future. Please let us know if there are any issues and bugs.

Framework Overview

BEV Feature Generation

MonoTAKD DEMO

• Detection in CAMERA perspective

• Detection with CAMERA & BEV Side-By-Side

Performance

KITTI

AP_3D performance on the KITTI test set for the car category.

	Teacher	TA	Student	Easy	Moderate	Hard
MonoTAKD	SECOND	CaDDN	model	27.91	19.43	16.51
MonoTAKD_Raw	SECOND	CaDDN	model	29.86	21.26	18.27

Nuscenes

	NDS	mAP
BEVFormer-R50 + TAKD	49.0	39.2
BEVFormer-R101 + TAKD	55.8	45.1
BEVDepth-R50 + TAKD	53.7	43.0
BEVDepth-R101 + TAKD	56.4	46.6

Set Up MonoTAKD

Installation

Please follow INSTALL to install MonoTAKD.

Getting Started

Please follow GETTING_START to train or evaluate the models.

Visualize Detection

Please follow VISUALIZE_DETECTION to draw detection bounding boxes onto 3D perspective view and BEV view.

Upload test set to KITTI Benchmark

Please follow KITTI_TEST_UPLOAD_GUIDELINES to upload to KITTI Benchmark for evaluation.