Reproduce “CHARM3R: Towards Unseen Camera Height Robust Monocular 3D Detector” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Abhinav Kumar, Yuliang Guo, Zhihao Zhang, Xinyu Huang, Liu Ren, Xiaoming Liu

CHARM3R introduces a monocular 3D object detector robust to varying camera heights by averaging geometrically-derived ground-based depth estimates with neural network regression-based depth estimates. This approach leverages their opposing error trends to improve 3D detection performance by over 45% on unseen camera heights compared to baselines.

Key insights

  • Depth estimation is the primary factor contributing to the significant performance degradation of state-of-the-art Mono3D models under varying ego camera heights.
  • Neural network-based depth regression exhibits a negative mean depth error trend (underestimation) with increasing camera height, while ground-based geometric depth estimation shows a positive trend (overestimation).
  • Averaging these two depth estimates, which have opposing error trends, effectively cancels out their individual biases, leading to more stable and accurate depth predictions for out-of-distribution camera heights.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.