ASL-DVS

This is a denoised, windowed derivative of the ASL-DVS event-camera dataset. Each row is a 1 second asynchronous event window from the original DAVIS240C recordings.

The original events are preserved in the source sensor coordinate system (240x180). Events are not converted to frames and are not cropped. For convenience, each row includes a recommended 128x128 crop location as metadata only.

Upstream Dataset Credit

The original ASL-DVS dataset was introduced in:

Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze and Yiannis Andreopoulos. "Graph-Based Object Classification for Neuromorphic Vision Sensing." Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

Useful upstream links:

• Original ASL-DVS repository maintained by the dataset creators
• ICCV 2019 paper
• DOI: 10.1109/ICCV.2019.00058

@inproceedings{Bi_2019_ICCV,
  author = {Bi, Yin and Chadha, Aaron and Abbas, Alhabib and Bourtsoulatze, Eirina and Andreopoulos, Yiannis},
  title = {Graph-Based Object Classification for Neuromorphic Vision Sensing},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2019}
}

Representation

Each row contains one denoised event window:

Column	Description
sample_id	Unique window id
recording_id	Source recording id
subject_id	Subject identifier
label	Lowercase ASL letter
label_id	Numeric class index
split	train or test
window_index	Window index within the recording
window_start_us	Window start relative to the recording start
num_events	Number of denoised events in the row
t	Event timestamps relative to the window start, in microseconds
x	Horizontal coordinate in [0, 239]
y	Vertical coordinate in [0, 179]
p	Event polarity
crop_128_left	Suggested 128x128 crop left coordinate
crop_128_top	Suggested 128x128 crop top coordinate

Preprocessing

• raw format: AER-DAT2.0 DAVIS240C;
• source resolution: 240x180;
• window duration: 1000000 microseconds;
• window overlap: 100000 microseconds (10%);
• stride: 900000 microseconds;
• denoise: equivalent to tonic.transforms.Denoise(filter_time=10000), applied once to each continuous recording before windowing;
• minimum events per exported window after denoise: 1000;
• crop metadata: computed from the denoised events in that row by selecting the 128x128 window with the highest event count.

The recommended crop is metadata. The event coordinates remain uncropped.

Visual Example

The following example uses a 300 ms excerpt from subject1/a.aedat, displayed on the full 240x180 sensor. The dataset rows are still 1 second long; the shorter excerpt is used only to make the hand shape easier to inspect in the dataset card. Blue and orange pixels represent the two event polarities.

Raw 300 ms events, full sensor	Denoised 300 ms events, full sensor

The dataset does not crop the events. The crop box below is only the suggested 128x128 region stored in crop_128_left and crop_128_top.

Classes

The dataset uses 24 static ASL letter classes:

a b c d e f g h i k l m n o p q r s t u v w x y

Letters j and z are not included because they require motion.

Split Policy

The default split is subject-independent:

Split	Subjects
train	subject1, subject2, subject3, subject4
test	subject5

The locally available raw subset does not contain subject5/g.aedat; therefore the test split lacks class g.

Dataset Summary

{
  "name": "ASL-DVS",
  "num_rows": 12146,
  "splits": {
    "train": 7227,
    "test": 4919
  },
  "classes": [
    "a",
    "b",
    "c",
    "d",
    "e",
    "f",
    "g",
    "h",
    "i",
    "k",
    "l",
    "m",
    "n",
    "o",
    "p",
    "q",
    "r",
    "s",
    "t",
    "u",
    "v",
    "w",
    "x",
    "y"
  ],
  "source_sensor_size": [
    240,
    180
  ],
  "window_duration_us": 1000000,
  "window_overlap_us": 100000,
  "window_stride_us": 900000,
  "denoise_filter_time_us": 10000,
  "min_events_per_window": 1000,
  "crop_metadata_size": [
    128,
    128
  ],
  "excluded_recordings": [
    {
      "recording_id": "subject2_k",
      "note": "excluded: anomalous sparse timeline; no valid denoised 1s windows"
    },
    {
      "recording_id": "subject2_xtrash",
      "note": "excluded: marked as trash"
    },
    {
      "recording_id": "subject3_x",
      "note": "excluded: anomalous sparse timeline; no valid denoised 1s windows"
    }
  ]
}

Limitations

This is a windowed derivative intended for event-camera and SNN experiments. It is not a general-purpose ASL understanding dataset and should not be used to claim recognition of full ASL vocabulary, grammar or continuous signing.

The upstream repository does not provide an explicit redistribution license in this project. Confirm redistribution terms before making derivative releases public.