2020-07-14 - Released patch v1.1 fixing some corrupt images - you will receive a link to download it if you already requested the data.
Mapillary Street-level Sequences (MSLS) is a large-scale long-term place recognition dataset that contains 1.6M street-level images.
- ⬇️ Download: https://www.mapillary.com/dataset/places (sample here)
- 📄 Paper: https://research.mapillary.com/publication/cvpr20c
- ️🧑⚖️ Code of Conduct
- 🗳️ Contributing / Pull Requests
We've included an implementation of a PyTorch Dataset in datasets/msls.py. It can be used for evaluation (returning database and query images) or for training (returning triplets). Check out the demo to understand its usage.
A standalone evaluation script is available for all tasks. It reads the predictions from a text file (example) and prints the metrics.
Here we show results of models consisting of a Resnet50 backbone followed by Generalized Mean Layer. The models are trained with either the standard triplet loss or the uncertainty-aware Bayesian triplet loss. All models are trained with standard hard negative mining on image resolution 224x224.
Results on test set (Miami, Athens, Buenos Aires, Stockholm, Bengaluru, Kampala):
| Loss | R@1 | R@5 | R@10 | R@20 | M@1 | M@5 | M@10 | M@20 |
|---|---|---|---|---|---|---|---|---|
| Triplet Loss | 0.372 | 0.522 | 0.582 | 0.636 | 0.372 | 0.261 | 0.234 | 0.228 |
| Bayesian Triplet Loss | 0.366 | 0.513 | 0.574 | 0.629 | 0.366 | 0.253 | 0.229 | 0.222 |
Results on validation set (San Francisco, Copenhagen)
| Loss | R@1 | R@5 | R@10 | R@20 | M@1 | M@5 | M@10 | M@20 |
|---|---|---|---|---|---|---|---|---|
| Triplet Loss | 0.623 | 0.780 | 0.830 | 0.859 | 0.623 | 0.432 | 0.380 | 0.372 |
| Bayesian Triplet Loss | 0.618 | 0.746 | 0.805 | 0.839 | 0.618 | 0.419 | 0.369 | 0.360 |
images_vol_X.zip: images, split into 6 parts for easier download.metadata.zip: a single zip archive containing the metadata.patch_vX.Y.zip: unzip any patches on top of the dataset to upgrade.
All the archives can be extracted in the same directory resulting in the following tree:
- train_val
city- query / database
- images/
key.jpg - seq_info.csv
- subtask_index.csv
- raw.csv
- postprocessed.csv
- images/
- query / database
- test
city- query / database
- images/
key.jpg - seq_info.csv
- subtask_index.csv
- images/
- query / database
The meta files include the following information:
-
raw.csv: raw data recorded during capture
- key
- lon
- lat
- ca
- captured_at
- pano
-
seq_info.csv: Sequence information
- key
- sequence_id
- frame_number
-
postprocessed.csv: Data derived from the raw images and metadata
- key
- utm (easting and northing)
- night
- control_panel
- view_direction (Forward, Backward, Sideways)
- unique_cluster
-
subtask_index.csv: Precomputed image indices for each subtask in order to evaluate models on (all, summer2winter, winter2summer, day2night, night2day, old2new, new2old)
This repository is MIT licensed.