ActiveLight-PBR: A Physics-Guided Synthetic Dataset for Line-Scan Pavement Inspection

# ActiveLight-PBR Dataset **Official Dataset for:** "Physics-Guided Inverse Rendering for Line-Scan Pavement Inspection: Breaking the Bas-Relief Ambiguity via Spatial-Illumination Priors" ## 1. Overview The ActiveLight-PBR dataset is a highly rigorously controlled, physically-based rendered (PBR) synthetic dataset designed specifically for line-scan pavement inspection. It provides perfectly decoupled ground truths for macro-illumination, micro-normal (3D relief), and surface albedo, offering an unprecedented testbed for solving the Bas-Relief Ambiguity in active lighting environments. ## 2. Directory Structure & Global Physical Priors ```text ActiveLight-PBR_Release/ ├── README.md <-- This file ├── Global_Left_Prior.npy <-- Pre-computed spatial-illumination prior (Left Camera) ├── Global_Right_Prior.npy <-- Pre-computed spatial-illumination prior (Right Camera) ├── Train/ <-- Training sequences │ ├── L-00034905-00000005/ │ │ ├── ..._synthetic_road.npy │ │ ├── ..._pure_shading.npy │ │ ├── ..._aligned_normal.npy │ │ ├── ..._aligned_albedo.npy │ │ ├── ..._road_spliced_mask_padded.npy │ │ └── ..._info.json │ └── ... └── Val/ <-- Validation sequences ``` ### Why are there only TWO Global Prior files? (Important Note for Reproducibility) In our Hierarchical Prior Injection Framework (HPIF), the physical light prior is designed to strictly capture the **macroscopic illumination gradient** caused by the rigid multi-LED layout, while deliberately ignoring micro-topological depth (Z=0) to mathematically prevent any geometric data leakage (i.e., the network cannot "cheat" by looking at true depth variations in the prior). Due to the strict rigid-body kinematics of the vehicle-mounted line-scan system (the relative pose between cameras and LED bars is perfectly static), the spatial illumination field is **globally constant** for a given camera viewpoint. Therefore, instead of duplicating prior maps for every sample, we extract the absolute physical coordinates and compute **two global, static spatial-illumination field tensors** (`Global_Left_Prior.npy` and `Global_Right_Prior.npy`). These 4-channel tensors contain the exact physical attenuation and anti-symmetric dominant light direction vectors corresponding to the left and right camera FOVs. ## 3. Sample File Modalities Each sample directory contains strictly 6 files, carefully curated for intrinsic decomposition evaluation: 1. `*_synthetic_road.npy`: The realistic rendering modality coupling albedo and illumination. 2. `*_pure_shading.npy`: The pure shading modality (albedo locked to a constant, e.g., 0.2), used for mechanistic probing. 3. `*_aligned_normal.npy`: Surface normal Ground Truth. 4. `*_aligned_albedo.npy`: Surface albedo Ground Truth. 5. `*_road_spliced_mask_padded.npy`: Boolean mask indicating valid asphalt regions for metric evaluation. 6. `*_info.json`: Optical and hardware metadata. ## 4. Usage When using the official GitHub repository, the `dataloader.py` will automatically fetch the corresponding `Global_Left_Prior.npy` or `Global_Right_Prior.npy` based on the sequence prefix (`L-` or `R-`) and inject it into the neural network, perfectly reproducing the physics-guided training process without requiring the proprietary low-level rendering engine. ## 5. License This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. It is freely available for academic and research purposes. Commercial use requires explicit authorization from RoadMainT Co., Ltd.