One potential shortcoming of GL3D is that images are captured within a short period of time, thus lack of illumination/weather/season variations. Although photometric data augmentation could be applied, we still seek for more realistic data to improve the learning models.
To this end, following SiaMAC [1], we have also generated geometric labels from public Internet tourism datasets to further increase the data diversity. Precisely, we download and extract the images from retrieval-SfM-120k.mat, then reconstruct each data by our 3D engine, and finally obtain 530 scenes (55,657 images) that we consider are well-constructed (> 80% images are registered).
The same protocols are defined for downloading the data. For dataset images:
| Sources | Data Name | Chunk Start | Chunk End | Disk | Descriptions |
|---|---|---|---|---|---|
| tourism | tourism_raw_imgs | 0 | 38 | 19G | Original images of tourism dataset |
| tourism | tourism_imgs | 0 | 29 | 15G | 1000x1000 undistorted images of tourism dataset |
For geometric labels:
| File Name | Data Name | Chunk Start | Chunk End | Disk | Task |
|---|---|---|---|---|---|
| geolabel/cameras.txt | tourism_cams | 0 | 0 | <0.1G | Common |
| img_kpts/<img_idx>.bin | tourism_kpts | 0 | 12 | 6.0G | Common |
| depths/<img_idx>.pfm | tourism_depths | 0 | 24 | 12G | Common |
| geolabel/corr.bin | tourism_corr | 0 | 9 | 4.5G | Local descriptor |
| geolabel/mask.bin | tourism_mask | 0 | 13 | 6.5G | Image retrieval |
| geolabel/common_track.txt | tourism_ct | 0 | 0 | <0.1G | Image retrieval |
| geolabel/mesh_overlap.txt | tourism_mo | 0 | 0 | <0.1G | Image retrieval |