Upload 9 files

.gitattributes

@@ -0,0 +1 @@
+MapItAnywhere/samples.pdf filter=lfs diff=lfs merge=lfs -text

MapItAnywhere/README.md

@@ -0,0 +1,41 @@
+# MIA Dataset
+
+## Description
+The MIA dataset is a bird's eye view (BEV) map prediction given first person view RGB images dataset. The dataset is part of the [MIA project](https://mapitanywhere.github.io/) and was curated using the [MIA data engine code](https://github.com/MapItAnywhere/MapItAnywhere) given first person view RGB images. The dataset has 1.2M data samples covering Pittsburgh, Los Angeles, San Francisco, Chicago, New York, and Houston.  
+
+## Format
+
+Each data sample has a unique ID given by Mapillary and is used to reference and associate attributes related to the sample throughout the dataset.
+
+**Each location has the following:**
+
+- `images` Directory containing all FPV images named as `<id>_undistorted.png`
+- `semantic_masks` npz files named as `<id>` containing semantic masks in the format of a single array `arr_0` with shape 224x224x8 where the 3rd dimension maps to classes as follows:
+	0. road
+	1. crossing
+	2. explicit_pedestrian
+	3. park (Unused by Mapper)
+	4. building
+	5. water (Unused by Mapper)
+	6. terrain
+	7. parking
+	8. train (Unused by Mapper)
+- `flood_masks` npz files named as `<id>` containing an observable region mask in the format of a single array `arr_0` with shape 224x224.
+- `image_points.parquet` dataframe containing all image points retrieved within the tiles encompassing the boundary.
+- `image_metadata.parquet` dataframe including metadata retrieved for each image point retrieved (After boundary filtering). The metadata retrieved is documented in the [Mapillary API](https://www.mapillary.com/developer/api-documentation#image)
+- `image_metadata_filtered.parquet` As above but only keeping filtered records
+- `image_metadata_filtered_processed.parquet` the final dataframe after FPV processing and spatial filtering and is the one that reflects what to expect in `images` directory.
+- `dump.json` a json file containing camera intrinsics and extrinsics for each image taken. Same format as [OrienterNet](https://github.com/facebookresearch/OrienterNet).
+
+In addition `split.json` is a file at the root that describes our training, validation, and testing splits.
+
+**Note** that throughout the pipeline, some data samples are unable to be processed fully due to API issues or processing limitations. Such data samples may have residues in dataframes or split files but may not have corresponding maps or flood masks. Thus, a valid data sample is defined as one that has a corresponding image, metadata record, semantic mask, and flood mask. The invalid data samples are less than 0.001% and will be cleaned up in later versions.
+
+
+## Licenses
+The FPVs were curated and processed from Mapillary and have the same [CC by SA license](https://creativecommons.org/licenses/by-sa/4.0/deed.en). These include all images files, parquet dataframes, and dump.json.
+The BEVs were curated and processed from OpenStreetMap and has the same [Open Data Commons Open Database  (ODbL)  License](https://opendatacommons.org/licenses/odbl/). These include all semantic masks and flood masks.
+The rest of the data is licensed under [CC by SA license](https://creativecommons.org/licenses/by-sa/4.0/deed.en)
+
+## Citation
+If you use this dataset please cite us and visit our [website](https://mapitanywhere.github.io/) for the latest citation information.

MapItAnywhere/samples.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e630d9171d5be9232d0f8e3e8e7c6a99c174edd92dcca752574a991bf6225e9b
+size 3048955