Lightweight Facial Landmark Detection in Thermal Images via Multi-Level Cross-Modal Knowledge Transfer
Qiyi Tong, Olivia Nocentini, Marta Lagomarsino, Kuanqi Cai, Marta Lorenzini, Arash Ajoudani
TL;DR
This work tackles facial landmark detection in thermal images by bridging the modality gap between RGB and thermal data through a decoupled, two-level framework called Multi-Level Cross-Modal Knowledge Distillation (MLCM-KD). The core innovation is Dual-Injected Knowledge Distillation (DIKD), a bidirectional cross-modal supervision mechanism that jointly injects RGB features into the thermal head (Forward Injection) and validates the thermal features within the RGB head (Reverse Injection), promoting modality-invariant, semantically aligned representations. The approach separates cross-modal transfer (KTL) from model compression (MCL), and employs a time-adaptive distillation decay to stabilize training while maintaining accuracy. Across public benchmarks, MLCM-KD achieves state-of-the-art performance with significantly reduced model complexity and real-time inference, including robust edge-device performance, demonstrating practical applicability for all-weather FLD tasks. The work also provides extensive ablations and analyses, highlighting the complementary roles of FI and RI and establishing a principled framework for cross-modal knowledge transfer that can extend to other sensing modalities and temporal analysis.
Abstract
Facial Landmark Detection (FLD) in thermal imagery is critical for applications in challenging lighting conditions, but it is hampered by the lack of rich visual cues. Conventional cross-modal solutions, like feature fusion or image translation from RGB data, are often computationally expensive or introduce structural artifacts, limiting their practical deployment. To address this, we propose Multi-Level Cross-Modal Knowledge Distillation (MLCM-KD), a novel framework that decouples high-fidelity RGB-to-thermal knowledge transfer from model compression to create both accurate and efficient thermal FLD models. A central challenge during knowledge transfer is the profound modality gap between RGB and thermal data, where traditional unidirectional distillation fails to enforce semantic consistency across disparate feature spaces. To overcome this, we introduce Dual-Injected Knowledge Distillation (DIKD), a bidirectional mechanism designed specifically for this task. DIKD establishes a connection between modalities: it not only guides the thermal student with rich RGB features but also validates the student's learned representations by feeding them back into the frozen teacher's prediction head. This closed-loop supervision forces the student to learn modality-invariant features that are semantically aligned with the teacher, ensuring a robust and profound knowledge transfer. Experiments show that our approach sets a new state-of-the-art on public thermal FLD benchmarks, notably outperforming previous methods while drastically reducing computational overhead.