Detection of object boundary inconsistencies between 2D-3D conversion results and depth maps
- Author: Stanislav Dolganov
- Supervisor: dr. Dmitriy Vatolin
Introduction
The creation of S3D movies by converting 2D captured footage requires the high-quality depth maps. In case of inaccurate depth maps’ usage during the conversion different types of artifacts can occur. Such artifacts can make the viewing experience significantly worse even if it appears only in the background.
Our method extracts information about motion in the scene and detects conversion-related discrepancies between motion strength and depth.
An example of a scene containing a defect found by the proposed method: the foreground object was merged with the background.
Proposed method
Algorithm scheme
The disparity map estimation step may be skipped if the original depth map is available.
Experiments
The algorithm was tested on 39 converted stereoscopic feature movies. As a result, 125 scenes with this artifact were found.
The number of scenes with notable artifacts found during the analysis of feature movies are illustrated below.
The examples of detected scenes are listed below. In these scenes, foreground objects merge with the background.
-
MSU Benchmark Collection
- MSU Video Upscalers Benchmark 2022
- MSU Video Deblurring Benchmark 2022
- MSU Video Frame Interpolation Benchmark 2022
- MSU HDR Video Reconstruction Benchmark 2022
- MSU Super-Resolution for Video Compression Benchmark 2022
- MSU No-Reference Video Quality Metrics Benchmark 2022
- MSU Full-Reference Video Quality Metrics Benchmark 2022
- MSU Video Alignment and Retrieval Benchmark
- MSU Mobile Video Codecs Benchmark 2021
- MSU Video Super-Resolution Benchmark
- MSU Shot Boundary Detection Benchmark 2020
- MSU Deinterlacer Benchmark
- The VideoMatting Project
- Video Completion
- Codecs Comparisons & Optimization
- VQMT
- MSU Datasets Collection
- Metrics Research
- Video Quality Measurement Tool 3D
- Video Filters
- Other Projects