FIMKD: Feature-Implicit Mapping Knowledge Distillation for RGB-D Indoor Scene Semantic Segmentation

Depth images are often used to improve the geometric understanding of scenes owing to their intuitive distance properties. Although there have been significant advancements in semantic segmentation tasks using red–green–blue-depth (RGB-D) images, the complexity of existing methods remains high. Furthermore, the requirement for high-quality depth images increases the model inference time, which limits the practicality of these methods. To address this issue, we propose a feature-implicit mapping knowledge distillation (FIMKD) method and a cross-modal knowledge distillation (KD) architecture to leverage deep modal information for training and reduce the model dependence on this information during inference. The approach comprises two networks: FIMKD-T, a teacher network that uses RGB-D data, and FIMKD-S, a student network that uses only RGB data. FIMKD-T extracts high-frequency information using the depth modality and compensates for the loss of RGB details due to a reduction in resolution during feature extraction by the high-frequency feature enhancement module, thereby enhancing the geometric perception of semantic features. In contrast, the FIMKD-S network does not employ deep learning techniques; instead, it uses a nonlearning approach to extract high-frequency information. To enable the FIMKD-S network to learn deep features, we propose a feature-implicit mapping KD for feature distillation. This mapping technique maps the features in channel and space to a low-dimensional hidden layer, which helps to avoid inefficient single-pattern student learning. We evaluated the proposed FIMKD-S* (FIMKD-S with KD) on the NYUv2 and SUN-RGBD datasets. The results demonstrate that both FIMKD-T and FIMKD-S* achieve state-of-the-art performance. Furthermore, FIMKD-S* provides the best performance balance.