Multi-Scale Spatio-Temporal Feature Extraction and Depth Estimation from Sequences by Ordinal Classification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Depth estimation is crucial for 3D computer vision applications.
• Existing methods face challenges in accurately predicting depth maps from sequences.

Purpose of the Study:

• To investigate the efficacy of deep learning networks in depth estimation.
• To improve depth map prediction accuracy by learning multi-scale spatio-temporal features.

Main Methods:

• Recasting depth estimation as an ordinal classification task instead of regression.
• Designing an encoder-decoder network incorporating multi-scale strategies.
• Utilizing Convolutional Long Short-Term Memory (ConvLSTM) for spatio-temporal feature extraction.

Main Results:

• The proposed method achieved nearly 10% improvement in error metrics.
• Accuracy metrics saw an improvement of up to 2%.
• Extracting spatio-temporal features significantly boosted performance.

Conclusions:

• Deep learning, particularly with spatio-temporal features, enhances depth estimation accuracy.
• Future work aims to develop self-supervised methods to reduce reliance on labeled data.