Dataset Card for sama-drives-california

Dataset Summary

This is an object detection dataset (bounding boxes and polygons) of 25 136 frames (848x480 pixels) taken by a dashboard video camera of a car driving in California. The frames were captured at 1 FPS, and hence the entire footage covers over 7 hours of driving. All but 110 frames contain at least one annotated object (25 026) of interest.

Dataset Structure

Data Instances

The dataset is saved according to the bdd100k format described here (no affiliation with Sama). Frames are named according to the original video they are from, along with the sequence index in that video (1-indexed): videoNumber-frameIndex.jpg
(e.g., 099-002.jpg for the second frame of the 99th video)

label:ids are used to denote unique objects, such as a specific vehicle, throughout an entire video, but not across videos. The first digits of a label:id denote what video it is from (e.g., the id 53002 comes from video 53).

Frames were taken from videos that were recorded in a continuous sequence without any time gap in between videos. However, some videos were not included
in the final dataset either because they contained sensitive information or because they were part of a long sequence when the car was parked and facing a scene of no interest.

The labelling format and different classes supported are described in the section Data Fields below.

Sample annotation:

{
  "name": "001-019.jpg",
  "attributes": {"weather": "Sunny", "timeofday": "Day"},
  "labels":
  [
    {"category": "Drivable Space", "attributes": {"occluded": true}, "manualShape": true, "manualAttributes": true, "id": 1001, "poly2d": [{"vertices": [[369, 296], [370, 276], [389, 277], [432, 278], [494, 279], [504, 266], [563, 262], [590, 270], [656, 271], [705, 276], [776, 270], [847, 274], [847, 337], [847, 419], [766, 408], [681, 402], [626, 400], [550, 393], [507, 391], [426, 390], [321, 387], [242, 394], [206, 402], [170, 402], [135, 399], [72, 405], [29, 413], [0, 418], [0, 259], [66, 259], [91, 267], [154, 265], [126, 280], [145, 288], [188, 284], [155, 265], [187, 265], [225, 263], [309, 260], [301, 271], [345, 272], [370, 276], [369, 296], [306, 300], [225, 300], [226, 312], [309, 334], [416, 353], [552, 373], [635, 375], [669, 365], [666, 343], [654, 338], [542, 313]], "closed": true, "filled": true}], "box2d": {"x1": 0, "y1": 259, "x2": 847, "y2": 419}},
    {"category": "Vehicles | Truck", "attributes": {"occluded": true}, "manualShape": true, "manualAttributes": true, "id": 1041, "poly2d": [{"vertices": [[708, 247], [692, 247], [688, 251], [687, 258], [687, 265], [709, 265], [714, 265], [713, 255]], "closed": true, "filled": true}], "box2d": {"x1": 687, "y1": 247, "x2": 714, "y2": 265}},
    {"category": "Vehicles | Truck", "attributes": {"occluded": true}, "manualShape": true, "manualAttributes": true, "id": 1043, "poly2d": [{"vertices": [[468, 238], [486, 251], [494, 253], [500, 257], [507, 258], [515, 262], [527, 267], [530, 278], [531, 293], [503, 300], [482, 299], [425, 291], [426, 296], [415, 298], [409, 291], [391, 288], [390, 299], [375, 300], [369, 289], [353, 284], [354, 254], [409, 256], [424, 238]], "closed": true, "filled": true}], "box2d": {"x1": 353, "y1": 238, "x2": 531, "y2": 300}},
    {"category": "Vehicles | Car", "attributes": {"occluded": true}, "manualShape": true, "manualAttributes": true, "id": 1044, "poly2d": [{"vertices": [[560, 256], [539, 253], [541, 257], [553, 264], [561, 271], [563, 288], [568, 288], [584, 290], [596, 288], [599, 277], [595, 271], [589, 267], [577, 264], [570, 260]], "closed": true, "filled": true}], "box2d": {"x1": 539, "y1": 253, "x2": 599, "y2": 290}},
    {"category": "Vehicles | Car", "attributes": {"occluded": true}, "manualShape": true, "manualAttributes": true, "id": 1045, "poly2d": [{"vertices": [[507, 246], [499, 247], [495, 248], [506, 255], [523, 262], [526, 270], [532, 281], [530, 295], [547, 296], [565, 294], [562, 271], [551, 261], [537, 254], [519, 251]], "closed": true, "filled": true}], "box2d": {"x1": 495, "y1": 246, "x2": 565, "y2": 296}},
    {"category": "Vehicles | Car", "attributes": {"occluded": false, "drivingConditions": "Light Traffic"}, "manualShape": true, "manualAttributes": true, "id": 1046, "poly2d": [{"vertices": [[30, 249], [14, 249], [9, 256], [8, 262], [10, 271], [13, 271], [13, 269], [24, 269], [24, 271], [30, 271], [32, 268], [36, 268], [38, 271], [41, 269], [41, 263], [40, 256], [37, 252], [34, 250]], "closed": true, "filled": true}], "box2d": {"x1": 8, "y1": 249, "x2": 41, "y2": 271}}
  ]
}

Data Fields

Each frame contains a label for timeofday and weather. Dusk, Dawn and Twilight all fall in the same timeofday category.

timeofday	weather
Day	Sunny
Night	Cloudy
Dusk/Dawn/Twilight	Rainy
	Snowy
	Other

Bounding boxes are provided for all objects as box2d. Vehicles, People and Areas are also identified with closed Polygons of the type poly2d. Lanes are available as Lines, that are denoted as open Polygons of the type poly2d. Traffic Lights and Traffic Signs are only available as Bounding Boxes.

Vehicles (Polygons)	People (Polygons)	Areas (Polygons)	Lanes (Lines)	Traffic (Bounding Boxes)
Car	Pedestrians	Drivable Space	Current Lane	Traffic Lights
Truck			Alternate Lane	Traffic Signs
Van			Opposite Lane
SUV
Bus
Other LV
Bicycles
Motorbikes

The objects above can each be occluded (true) or not (false).

Vehicles also have a label called drivingConditions that denotes the amount of vehicle traffic they are facing. Note that this label is not always present.

drivingConditions (for Vehicles)
Light Traffic
Moderate Traffic
Heavy Traffic

Lanes also contain a laneChange label. Note that this label is not always present.

laneChange (for Lanes)
Current
Alternate
Opposite

Visualize Dataset

To visualize the dataset on the FiftyOne app, download and unzip the following zip file (2.3GB).

import fiftyone as fo

# <dataset_dir>/
#    labels.json
#    data/
#        001-001.jpg
#        001-002.jpg
#        ...

name = "sama-drives-california"
dataset_dir = "/path/to/dataset"

# Create the dataset
dataset = fo.Dataset.from_dir(
    dataset_dir=dataset_dir,
    dataset_type=fo.types.BDDDataset,
    name=name
)

Dataset in Video Format

This dataset is also available as a video dataset with FiftyOne style label format. You can download a zipped file of the dataset (videos and fiftyone labels) here (1.1GB).

import fiftyone as fo

# <video_dataset_dir>/
#    frames.json
#    metadata.json
#    samples.json
#    data/
#        001.mp4
#        002.mp4
#        ...

name = "sama-drives-california-videos"
dataset_dir = "/path/to/videos-dataset"

# Create the dataset
dataset = fo.Dataset.from_dir(
    dataset_dir=dataset_dir,
    dataset_type=fo.types.FiftyOneDataset,
    name=name
)

Annotations

The dataset was annotated by a team of Sama Associates. They were instructed to annotate all objects of the classes described in the section Data Fields above with the following details:

• Ignore objects under 10 pixels in width or height.
• Annotate with a pixel tolerance of 2 pixels.
• For motorized vehicles, include the mirrors but do not include the antennas.
• For bicycles, include the cyclist.
• For motorbikes, include the rider.
• For traffic lights, place the bounding box around the light fixture but not the pole.
• For traffic signs, do not include the pole or structure.

Personal and Sensitive Information

All personal and sensitive information has been removed. Vehicle license plates and faces are blurred.

Other Known Limitations

Objects of interest that were smaller than 10 pixels in width or height were not annotated.

Licensing Information

(CC BY 4.0) [https://creativecommons.org/licenses/by/4.0/]