Learnability-Driven Submodular Optimization for
Active Roadside 3D Detection

Ruiyu Mao¹   ·   Baoming Zhang¹   ·   Nicholas Ruozzi¹   ·   Yunhui Guo¹

¹The University of Texas at Dallas
CVPR 2026

This repository contains the official implementation of LH3D — a learnability-driven active learning framework for vision-based roadside 3D object detection. Built on top of BEVHeight, LH3D selects the most informative training samples under a fixed annotation budget using a three-stage submodular selection strategy driven by depth learnability, spatial diversity, and geometric similarity.

---

Getting Started

Installation

See docs/install.md for environment setup.

Dataset Preparation

See docs/prepare_dataset.md for DAIR-V2X-I and Rope3D setup.

Training with LH3D

# DAIR-V2X-I
python exps/dair-v2x/bev_height_lss_r50_864_1536_128x128_active.py \
  --al_enabled \
  --al_method lh3d \
  --al_init_size 500 \
  --al_query_size 120 \
  --al_rounds 10 \
  --al_epochs_per_round 5 \
  --al_max_objects 32000 \

# Rope3D
python exps/rope3d/bev_height_lss_r50_864_1536_128x128_active.py \
  --al_enabled \
  --al_method lh3d \
  --al_init_size 500 \
  --al_query_size 120 \
  --al_rounds 10 \
  --al_epochs_per_round 5 \
  --al_max_objects 32000 \

---

Acknowledgment

This project builds on the following works:

• BEVHeight — roadside 3D detection backbone (CVPR 2023)
• BEVDepth — LSS-based depth estimation
• DAIR-V2X — dataset and evaluation toolkit
• pypcd — point cloud utilities

---

Incoming

• Release the pretrained models

---

Citation

If you find this work useful, please cite our paper:

@article{mao2026learnability,
  title={Learnability-Driven Submodular Optimization for Active Roadside 3D Detection},
  author={Mao, Ruiyu and Zhang, Baoming and Ruozzi, Nicholas and Guo, Yunhui},
  journal={arXiv preprint arXiv:2601.01695},
  year={2026}
}