Multi-View Material Classification on a Conveyor Belt: RGB–LWIR Video Dataset

22 unsynchronised RGB and LWIR (FLIR) video pairs of household waste objects (glass, metal, paper, plastic) on a conveyor belt, with 901 manually-labelled per-object tracklets, RGB-to-thermal homography matrices, precomputed SuperPoint–SuperGlue frame-matching CSVs, and a frozen Mask R-CNN detector (ResNet-50-FPN backbone) trained on the ThermalRGBTrash segmentation pool. These tracklets are divided into a 19-experiment cross-validation development set (550 tracklets) and a separate 3-experiment held-out mixed-class test set (351 tracklets). Optional convenience tier includes precomputed OC-SORT tracklets and frozen DINOv2 ViT-L/14 feature caches. Companion data to:"Software-Only Registration and Cross-Spectral Classification of Unsynchronised RGB–LWIR Video: A Multisensor Benchmark for Conveyor-Based Waste Sorting" (Akdemir & Yüksel, MDPI Sensors, 2026). Code repository: https://github.com/akde/mwc-rgbt-waste-sorting The pretrained Detectron2 weights in det2_model/ were trained on the ThermalRGBTrash dataset; ThermalRGBTrash has its own license, which governs redistribution of those weights (see det2_model/PROVENANCE.md inside the archive). Version 1.1.0. Adds three mixed-class test experiments (0855, 0861, 0864), each containing all four material classes, contributing 351 labeled tracklets (glass 90, metal 89, paper 70, plastic 102). The dataset now spans 22 experiments and 901 tracklets. These three experiments form the paper's held-out mixed-class test set (evaluated once; never used for cross-validation training or model selection). The 19 cross-validation experiments and the frozen detector are unchanged from v1.0.0.