zyg11 / admap

ADMap: Anti-disturbance framework for reconstructing online vectorized HD map

ADMap: Anti-disturbance framework for reconstructing online vectorized HD map

Due to confidentiality rules, the open source code is extra replicated. If it doesn't work correctly, ask a question in the issue and we'll fix the problem as soon as possible.

New

• 2024.1.22 create README.md

Abstract

In the field of autonomous driving, online high-definition (HD) map reconstruction is crucial for planning tasks. Recent research has developed several high-performance HD map reconstruction models to meet this necessity. However, the point sequences within the instance vectors may be jittery or jagged due to prediction bias, which can impact subsequent tasks. Therefore, this paper proposes the Anti-disturbance Map (ADMap) framework. To mitigate point-order jitter, the framework consists of three modules: Multi-Scale Perception Neck, Instance Interactive Attention (IIA), and Vector Direction Difference Loss (VDDL). By exploring the point-order relationships between and within instances in a cascading manner, the model can monitor the point-order prediction process more effectively. We confirmed the validity of ADMap in both nuScenes and Argoverse2, demonstrating its excellent performance.

paper

Main Result

nuScenes val

Method	Backbone	$AP_{div}$	$AP_{ped}$	$AP_{bou}$	mAP	FPS
MapTR	R50	51.5	46.3	53.1	50.3	15.1
ADMap	R50	56.2	49.4	57.9	54.5	14.8
MapTR	R50 & SECOND	55.9	62.3	69.3	62.5	6.0
ADMap	R50 & SECOND	66.6	63.3	74.0	68.0	5.8
MapTRv2	R50 & SECOND	65.6	66.5	74.8	69.0	5.8
ADMapv2	R50 & SECOND	67.9	68.5	74.5	70.3	6.1

• FPS is measured on NVIDIA RTX3090 GPU with batch size of 1.

Argoverse2 val

Method	Backbone	$AP_{div}$	$AP_{ped}$	$AP_{bou}$	mAP	FPS
MapTR	R50	65.5	56.6	61.8	61.3	14.8
ADMap	R50	68.9	60.3	64.9	64.7	14.2
MapTRv2	R50 & SECOND	62.9	72.1	67.1	67.4	12.0
ADMapv2	R50 & SECOND	72.4	64.5	68.9	68.7	13.9

• FPS is measured on NVIDIA RTX3090 GPU with batch size of 1.

Visualization results

nuScenes Visualization

Argoverse2 Visualization

Acknowlegement

We sincerely thank the authors of MapTR for open sourcing their methods.