Monocular depth estimation (MDE) has emerged as a powerful technique for extracting scene depth from a single image, particularly in the context of computational imaging. Conventional MDE methods based on RGB images often degrade under varying illuminations. To overcome this, an end-to-end framework is developed that leverages the illumination-invariant properties of infrared images for accurate depth estimation. Specifically, a multi-task UNet architecture was designed to perform gradient extraction, semantic segmentation, and texture reconstruction from infrared RAW images. To strengthen structural learning, a Radiation Field Gradient Guidance (RGG) module was incorporated, enabling edge-aware attention mechanisms. The gradients, semantics, and textures were mapped to the Saturation (S), Hue (H), and Value (V) channels in the HSV color space, subsequently converted into an RGB format for input into the depth estimation network. Additionally, a sky mask loss was introduced during training to mitigate the influence of ambiguous sky regions. Experimental validation on a custom infrared dataset demonstrated high accuracy, achieving a of 0.976. These results confirm that integrating radiation field gradient guidance and semantic priors in HSV space significantly enhances depth estimation performance for infrared imagery.
Monocular depth estimation offers cost-effective 3D perception, while thermal infrared sensors show high robustness under varying illumination conditions. Their combination is crucial for day and night 3D perception, yet it faces two key challenges: (1) inadequate processing of 14-bit RAW data, leading to loss of subtle infrared information, and (2) the resolution limitation of current infrared datasets (existing 640×512 sensor collections restrict the camera’s ability to capture critical scene details). To this end, in this paper, we introduce the IRSL dataset, which integrates infrared, stereo RGB, and LiDAR modalities. Data were acquired using a high-resolution thermal camera (1280×1024 pixels), synchronized stereo cameras, and LiDAR (500K points per frame). The dataset aims to enhance the precision and robustness of depth estimation in both daytime and nighttime environments. Secondly, this paper proposes Radiation Field Detail Enhancement (RFDE), a novel method for enhancing infrared radiation field details in RAW data. This method performs gray mapping based on radiation field intensity and spatial context awareness and uses a bilateral filter to enhance the details in the image. Experimental results show that the proposed RFDE method fully leverages the advantages of high-resolution infrared images, effectively improving the quantitative and qualitative performance of the MDE model (δ1 = 0.964, SqRel = 0.441).