This dataset contains 630 defocus image sequences and corresponding ground-truth with 1920x1080 resolution captured from 7 scenens ("Meetingroom"scene, "Livingroom"scene, "Bathroom"scene, "Street"scene, "Furniture" scene, "Machine" scene and "Industrialparts"scene). At present, for each scene, we provide 81 multi-focus images at the first frame with 240x135 resolution for download. The complete dataset will be released upon publication.
The source codes for our extensive validation experiments on depth estimation, 3D reconstruction and novel view synthesis tasks are now available. Please refer to the Code folder.
- Python >= 3.7
- PyTorch >= 1.4.0 is recommended
- opencv-python = 4.5.1
- matplotlib
- tensorboard
- pytorch_msssim
We release all the depth estimation, 3D reconstruction and novel view synthesis comparison results with 240x135 resolution between MVMFI and conventional multi-view video. For each scene, we reconstruct 81 dynamic depth map videos, 1 dynamic 3D point cloud video and synthesize 120 dynamic novel view videos. Please refer to the Results folder. The complete comparison results with 1920x1080 resolution will be released upon publication.
This dataset and results cannot be used for commercial purposes. For non-commercial research or educational purposes, please inform us for our permission via yangyou AT hust DOT edu DOT cn.