CA-ALIKED Dataset Subset: A Curated Multi-View Vehicle Sequence Collection from 3DRealCar Dataset

This dataset contains the curated subset of the 3DRealCar dataset used in the bachelor thesis "Accurate and Robust Localization of Landmarks on a Vehicle" (Adam Běhoun, Brno University of Technology, Faculty of Information Technology, 2026). It supports the development and evaluation of CA-ALIKED, a domain-adapted local feature extractor for reflective automotive surfaces. Contents: - train/ (115 unmasked sequences, 8,379 images) — used for self-supervised fine-tuning of the ALIKED detection head via multi-model Structure-from-Motion consensus - eval-imc/ (7 masked sequences) — used for IMC pose-estimation evaluation - eval-geo/ (5 masked sequences) — used for geometric stability analysis using Track Length and Reprojection Error A machine-readable manifest (dataset_splits.json) describes every sequence and frame in the archive. The data is derived from the 3DRealCar dataset (Du et al., ICCV 2025), which is distributed under the Apache 2.0 license and permits redistribution with attribution. The original 3DRealCar dataset and toolkit are available at https://github.com/xiaobiaodu/3DRealCar_Toolkit.