Deep Visual Discomfort Predictor for Stereoscopic 3D Images.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Perception Science

Background:

• Traditional stereoscopic 3D (S3D) visual discomfort prediction (VDP) relies on handcrafted features.
• These features are based on visual perception models and natural depth statistics.
• Deep learning offers a potential for improved VDP by automatically learning predictive features.

Purpose of the Study:

• To develop an advanced deep learning model for S3D visual discomfort prediction.
• To improve the accuracy and performance of VDP algorithms.
• To address the challenge of limited labeled data for training deep models.

Main Methods:

• Developed Deep Visual Discomfort Predictor (DeepVDP), a convolutional neural network (CNN) based model.
• Employed a patch-based CNN training strategy with two sequential steps.
• Utilized proxy ground-truth labels from an existing algorithm (3D-VDP) for initial training.
• Aggregated local features into global representations and fine-tuned with subjective discomfort scores.

Main Results:

• The DeepVDP model achieved state-of-the-art performance in predicting S3D visual discomfort.
• The patch-based training approach effectively leveraged available data.
• Learned features demonstrated high predictive power for experienced visual discomfort.
• The model successfully aggregated local abstractions into globally relevant features.

Conclusions:

• Deep learning, specifically CNNs, can significantly advance S3D visual discomfort prediction.
• The proposed patch-based training methodology is effective for deep VDP models.
• DeepVDP offers a robust and accurate solution for predicting visual discomfort in stereoscopic 3D content.