Car Surface Highlight Dataset

This dataset is the first high-gloss car dataset that includes natural lighting scenes. Its creation process uses a paired acquisition scheme of 'original high-gloss image + polarized de-gloss image': first, the original high-gloss image of the car is captured using a system fixed on a tripod; then the polarizer in the system is rotated to obtain the corresponding reflection-free image. Each image pair is captured in this sequence, resulting in 842 pairs of high-resolution images. For ease of model training, the image resolution is unified at 384×256. The acquisition equipment consists of a digital camera and a high-extinction-ratio circular polarizer. The use of a tripod effectively reduces positional offset errors during image pairing. In the data processing stage, each original high-gloss image is paired with a pixel-level gloss mask, accurately marking the location and range of gloss areas, without involving complex data augmentation or synthesis, ensuring the objectivity of real scene data.The dataset's spatial information covers outdoor environments with direct sunlight, parking lots illuminated by streetlights at night, and underground garage LED lighting scenarios. It includes various vehicle types such as sedans, SUVs, and trucks, as well as multiple car surface materials like metallic paint, matte finishes, and glass. It fully captures complex real-world phenomena such as multi-directional anisotropic highlights on metallic car bodies and highlight-reflection mixes on car windows. The time dimension is not restricted, focusing on coverage under different lighting conditions. The spatial resolution is based on high-resolution images, meeting the pixel-level highlight detection and removal algorithm evaluation needs. The tabular data in the dataset (Table 1) mainly records key comparative features of existing public datasets (such as SHIQ, SSHR, PSD, NSH, etc.), with row labels listing the dataset names and column labels including lighting conditions, whether the data is from real scenes, and highlight complexity-related indicators. The Structural Similarity Index (SSIM) is a unitless value used to quantify the structural differences between highlight images and reflection-free images.The dataset has no missing data. All 842 pairs of images are fully provided with the original specular highlight images, non-specular reflection images, and pixel-level highlight masks, ensuring good data completeness. Regarding errors, they may mainly arise from slight fluctuations in lighting conditions in real-world scenes (such as instantaneous changes in outdoor sunlight), but these have been minimized through fixed acquisition equipment and standardized collection procedures, with no systematic errors introduced. Image pairing errors have been significantly reduced by using a tripod and are within acceptable limits that do not affect algorithm evaluation. Each data file is centered around an image pair and contains three associated files: the original specular highlight image (recording the vehicle's highlight performance under target lighting conditions), the polarization-based specular-free image (a reference image obtained by removing specular reflection through polarization imaging), and the pixel-level highlight mask (a single-channel image marking the pixel-level locations of highlights). All files are stored in standard image formats, without obscure formats, and can be directly accessed and processed using mainstream image viewing software (such as Adobe Photoshop, IrfanView) and computer vision development tools (such as OpenCV, PyTorch). Compared to existing synthetic datasets or lab-controlled lighting datasets, the specular highlight images in this dataset exhibit greater chromatic shifts and higher texture complexity, providing a more realistic benchmark for evaluating highlight removal algorithms.