In the realm of monocular depth estimation, a persistent challenge arises: large models, while demonstrating robust generalization capabilities, pose significant deployment difficulties. Conversely, traditional lightweight models often fall short in terms of generalization. Addressing this dilemma, DepthART pioneers a viable approach that preserves cross-scenario depth estimation generalization abilities through a model boasting merely around 6 million parameters.
This research unveils a bias-resistant data sampling technique aimed at balancing the distribution of multi-source training data. This prevents small models from succumbing to biases and incorporates depth priors into these models via distillation methods. Furthermore, a camera-conditional fine-tuning strategy is devised to maintain the general geometric capabilities already acquired by the small model, facilitating low-cost recovery of true scale.
The DepthART-S model, with only a quarter of the parameters of Depth Anything V2-S, markedly enhances inference speed while upholding high generalization accuracy and supporting diverse deployment methods. A subsequent iteration, based on MobileNetV4, substitutes custom operators with standard ones, substantially reducing computational demands and the challenges associated with cross-hardware porting.
Currently, DepthART encompasses models of varying parameter scales, catering to different hardware constraints and propelling the advancement of lightweight monocular depth estimation technology. Relevant achievements have been accepted for presentation at ACM Multimedia 2026.