Monocular Depth Estimation with Augmented Ordinal Depth Relationships.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised depth estimation from monocular images typically requires extensive metric ground truth data.
• Acquiring accurate metric depths from stereo videos is often challenging due to missing camera parameters.

Purpose of the Study:

• To enhance metric depth estimation from single monocular images by leveraging relative depth cues from stereo videos.
• To introduce a new dataset and methodology for improved depth prediction.

Main Methods:

• Developed a new "Relative Depth in Stereo" (RDIS) dataset with dense relative depth labels.
• Pretrained a ResNet model on RDIS, then finetuned it on RGB-D datasets using a classification-based depth estimation approach.
• Introduced an information gain loss function utilizing prediction confidence (probability distribution).

Main Results:

• Achieved state-of-the-art performance on both indoor and outdoor benchmark RGB-D datasets.
• Demonstrated the effectiveness of relative depth cues for metric depth estimation.
• The classification-based formulation provided depth prediction confidence.

Conclusions:

• Relative depth from stereo videos is a valuable and accessible cue for improving monocular depth estimation.
• The proposed method, incorporating RDIS and an information gain loss, offers a robust approach to metric depth estimation.
• The technique successfully overcomes limitations of traditional supervised learning methods.