[Paper Review] LmPT: Conditional Point Transformer for Anatomical Landmark Detection on 3D Point Clouds

Matteo Bastico, Pierre Onghena+6 more|arXiv (Cornell University)|Feb 2, 2026

Forensic Anthropology and Bioarchaeology Studies0 citations

TL;DR

LmPT introduces a conditional point transformer with FiLM-based input conditioning to detect anatomical landmarks on cross-species 3D femur point clouds, achieving state-of-the-art accuracy on human and dog datasets and enabling translational cross-species learning.

ABSTRACT

Accurate identification of anatomical landmarks is crucial for various medical applications. Traditional manual landmarking is time-consuming and prone to inter-observer variability, while rule-based methods are often tailored to specific geometries or limited sets of landmarks. In recent years, anatomical surfaces have been effectively represented as point clouds, which are lightweight structures composed of spatial coordinates. Following this strategy and to overcome the limitations of existing landmarking techniques, we propose Landmark Point Transformer (LmPT), a method for automatic anatomical landmark detection on point clouds that can leverage homologous bones from different species for translational research. The LmPT model incorporates a conditioning mechanism that enables adaptability to different input types to conduct cross-species learning. We focus the evaluation of our approach on femoral landmarking using both human and newly annotated dog femurs, demonstrating its generalization and effectiveness across species. The code and dog femur dataset will be publicly available at: https://github.com/Pierreoo/LandmarkPointTransformer.

Motivation & Objective

• Motivate accurate, scalable automatic landmarking on 3D point clouds to reduce manual labelling variability.
• Develop a cross-species capable landmark detector that leverages homologous bones for translational research.
• Propose a conditional transformer model (LmPT) with FiLM modulation to adapt to different input types.
• Create and release a dog femur landmark dataset to enable translational cross-species evaluation.
• Demonstrate generalization across human and dog femurs with quantitative and qualitative analyses.

Proposed method

• Adopt a transformer-based encoder–decoder architecture (Point Transformer) for 3D point clouds.
• Introduce input-type conditioning via Feature-wise Linear Modulation (FiLM) applied to bottleneck encoder features.
• Train with a cross-entropy loss over labeled keypoints, ignoring unlabeled points.
• Normalize, sample to 8192 points, and augment with rotations, scaling, and coronal flips to improve cross-domain generalization.
• Evaluate using MAE and PCK across multiple thresholds, comparing against baselines including A&A and DGCNN.

Experimental results

Research questions

• RQ1 Can a conditional Point Transformer detect anatomical landmarks on 3D femur point clouds across species (human and dog)?
• RQ2 Does FiLM-based conditioning enable effective cross-species learning and improve landmark accuracy compared to single-species training?
• RQ3 How does LmPT-v2 compare to LmPT-v3 and other baselines in terms of landmark localization accuracy and convergence?
• RQ4 What is the impact of cross-species training on landmark detection performance for humans and dogs?
• RQ5 Can a newly annotated dog femur dataset support translational research and cross-species generalization?

Key findings

• LmPT with PTv2 backbone achieves state-of-the-art landmark MAE on human femurs and often surpasses expert-level annotations.
• PTv3’s attention-serialization strategy underperforms PTv2 in this task, indicating a trade-off between efficiency and localization precision.
• LmPT-v2 consistently outperforms baselines (A&A, DGCNN) on the human femur dataset for multiple landmarks.
• On the dog femur dataset, LmPT-v2 yields the lowest MAE and fastest convergence among evaluated methods.
• Cross-species training with FiLM conditioning improves human landmark MAE over single-species training, while dog MAE shows a nuanced response due to landmark set differences.
• Cross-species training achieves comparable or better performance at lower thresholds for both species, indicating effective cross-species feature sharing.