NVDS: Towards Efficient and Versatile Neural Stabilizer for Video Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Video depth estimation requires temporally consistent depth maps.
• Existing methods like fine-tuning single-image models are inefficient and lack robustness.
• Data-driven approaches need well-designed models and extensive video depth data.

Purpose of the Study:

• To introduce NVDS for stabilizing inconsistent depth from single-image models in a plug-and-play manner.
• To present the Video Depth in the Wild (VDW) dataset, the largest natural-scene video depth dataset.
• To establish a robust baseline and data foundation for learning-based video depth estimation.

Main Methods:

• NVDS stabilizes depth from various single-image models without retraining.
• A large-scale VDW dataset with 14,203 videos and over two million frames was created.
• A bidirectional inference strategy adaptively fuses forward and backward predictions for enhanced consistency.

Main Results:

• NVDS achieves significant improvements in depth estimation consistency, accuracy, and efficiency.
• The method demonstrates versatility by extending to video semantic segmentation and downstream tasks.
• Evaluations on VDW and public benchmarks confirm the effectiveness of the proposed approach.

Conclusions:

• NVDS offers an efficient and robust solution for video depth estimation.
• The VDW dataset provides a valuable resource for advancing video depth research.
• The proposed method and dataset serve as a strong foundation for future work in the field.