Objective, Absolute and Hue-aware Metrics for Intrinsic Image Decomposition on Real-World Scenes: A Proof of Concept

TL;DR

This paper tackles the challenge of evaluating intrinsic image decomposition (IID) on real-world scenes by eliminating reliance on subjective ground-truth judgments. It introduces an objective pipeline that derives albedo from hyperspectral reflectance and LiDAR intensity, yielding absolute, hue-aware IID assessments, and optionally densifies sparse albedo maps using a spectral similarity dictionary. In a laboratory proof-of-concept with a color board, the method achieves a CIEDE2000 error of $6.75$ and a luminance correlation of $0.981$ with ground truth, outperforming RGB-based and several IID baselines. The approach provides a practical framework for objective IID evaluation and dense albedo estimation, enabling more reliable benchmarking and potential use in model training, with future work extending to more complex scenes.

Abstract

Intrinsic image decomposition (IID) is the task of separating an image into albedo and shade. In real-world scenes, it is difficult to quantitatively assess IID quality due to the unavailability of ground truth. The existing method provides the relative reflection intensities based on human-judged annotations. However, these annotations have challenges in subjectivity, relative evaluation, and hue non-assessment. To address these, we propose a concept of quantitative evaluation with a calculated albedo from a hyperspectral imaging and light detection and ranging (LiDAR) intensity. Additionally, we introduce an optional albedo densification approach based on spectral similarity. This paper conducted a concept verification in a laboratory environment, and suggested the feasibility of an objective, absolute, and hue-aware assessment. (This paper is accepted by IEEE ICIP 2025. )

Figures (7)

• Figure 1: (a) The existing method relies on human-judged annotations of relative reflectance, suffering from subjectivity, relative evaluation, and hue non-assessment. (b) our method evaluates IID quality with calculated albedo from hyperspectral images and LiDAR intensity, achieving objectivity, absolute-value evaluation, and hue-awareness. Optionally, the calculated albedo can be densified by spectral similarity algorithms.
• Figure 2: (a) RGB image contains white line and cast shadows due to passive sensing. (b) LiDAR depth map dose not contain the white line and cast shadows. On the other hand, (c) LiDAR intensity contains white line while eliminating the cast shadows due to its active sensing.
• Figure 3: Illustration of the proposed albedo densification method based on hyperspectral images. Step (1) is pre-processing to extract pixels with albedo values and prepare dictionary. Step (2) to (4) constitute the densification process: (2) use spectral data without albedo as query, (3) calculate similarity, and (4) assign densified albedo values. The spectral signatures represent that of each pixel.
• Figure 4: Experimental setup. Hyperspectral camera, LiDAR, and color board are prepared to obtain experimental data. To illuminate the board, we use an artificial light source. Additionally, a T-shaped object is placed between the light source and the color board to create cast shadows.
• Figure 5: Comparison between (a) WHDR annotation and (b) our proposed evaluation, by plotting color differences within the color board image. WHDR relies on human-judged annotations, leading to subjectivity and discrete-value evaluation. In contrast, ours computes albedo from hyperspectral images and LiDAR intensity, providing objective and continuous-value evaluation. Furthermore, compared to the RGB image, our albedo plots align closely with the ground truth, forming an almost straight line.