coco-v2e-native-voxel5
Five-channel event voxel grids with YOLO object-detection annotations, prepared from an evbench benchmark directory. TIFF images and label files are preserved byte-for-byte. Raw event streams and original RGB images are not included.
Contents and splits
| Physical split | Samples | Boxes |
|---|---|---|
| train | 586,283 | 4,127,651 |
| val | 24,760 | 176,381 |
The YAML split mappings are:
trainusesimages/train.valusesimages/val.
When validation and test point to the same directory, they contain the same samples; there is no independent validation set. Select a separate validation subset from training data, grouped by original source image/sequence, before tuning a model.
Each uncompressed TAR shard contains paired members:
images/<split>/<sample-key>.tif
labels/<split>/<sample-key>.txt
Shards are sorted by sample key and keep each TIFF/label pair together. These are download-and-extract archives, rather than WebDataset-format samples. The Hub dataset viewer is disabled; load the extracted files with the training pipeline.
Representation and annotations
Each TIFF has five grayscale pages, one per temporal bin, in time order. Preserve
all five pages when decoding. The renderer produces uint8 signed-polarity (ON − OFF)
voxel counts over a 50.0 ms window, normalized by the 99th percentile of
nonzero absolute counts across the grid, clipped to [-1, 1], and encoded as
uint8(128 + 127 * value). A value of 128 denotes no net events.
Labels are UTF-8 text: one class_id center_x center_y width height row per object.
Coordinates are normalized to image dimensions; class IDs are zero-based.
Empty label files denote samples with no boxes.
| ID | Class |
|---|---|
| 0 | person |
| 1 | bicycle |
| 2 | car |
| 3 | motorcycle |
| 4 | airplane |
| 5 | bus |
| 6 | train |
| 7 | truck |
| 8 | boat |
| 9 | traffic light |
| 10 | fire hydrant |
| 11 | stop sign |
| 12 | parking meter |
| 13 | bench |
| 14 | bird |
| 15 | cat |
| 16 | dog |
| 17 | horse |
| 18 | sheep |
| 19 | cow |
| 20 | elephant |
| 21 | bear |
| 22 | zebra |
| 23 | giraffe |
| 24 | backpack |
| 25 | umbrella |
| 26 | handbag |
| 27 | tie |
| 28 | suitcase |
| 29 | frisbee |
| 30 | skis |
| 31 | snowboard |
| 32 | sports ball |
| 33 | kite |
| 34 | baseball bat |
| 35 | baseball glove |
| 36 | skateboard |
| 37 | surfboard |
| 38 | tennis racket |
| 39 | bottle |
| 40 | wine glass |
| 41 | cup |
| 42 | fork |
| 43 | knife |
| 44 | spoon |
| 45 | bowl |
| 46 | banana |
| 47 | apple |
| 48 | sandwich |
| 49 | orange |
| 50 | broccoli |
| 51 | carrot |
| 52 | hot dog |
| 53 | pizza |
| 54 | donut |
| 55 | cake |
| 56 | chair |
| 57 | couch |
| 58 | potted plant |
| 59 | bed |
| 60 | dining table |
| 61 | toilet |
| 62 | tv |
| 63 | laptop |
| 64 | mouse |
| 65 | remote |
| 66 | keyboard |
| 67 | cell phone |
| 68 | microwave |
| 69 | oven |
| 70 | toaster |
| 71 | sink |
| 72 | refrigerator |
| 73 | book |
| 74 | clock |
| 75 | vase |
| 76 | scissors |
| 77 | teddy bear |
| 78 | hair drier |
| 79 | toothbrush |
Provenance
meta.json preserves the frame build settings and original statistics.
The sequence metadata records source coco, modality v2e, preset calibrated, resolution 640 × 480, sequence duration 500 ms and label interval 100000 µs.
When available, sequence_meta.json preserves the upstream sequence export settings.
Machine-specific absolute paths are replaced with <local-path>/<basename>.
manifest.json records the actual packaged sample/box counts and shard hashes.
The build-time v2e_preset argument alone does not establish how upstream sequences
were simulated; consult the sequence metadata and generating code.
TODO before public release: add the generating code URL/commit, source-dataset citations and a citation for this derived dataset.
Download, verify and extract
pip install --upgrade huggingface_hub pyyaml
hf download evbench/coco-v2e-native-voxel5 --repo-type dataset --local-dir ./dataset
cd dataset
python extract.py
The extracted root contains data.yaml, images/ and labels/.
Pass the absolute path of data.yaml to evbench or Ultralytics. The release YAML
omits the machine-specific path. extract.py verifies all release checksums,
extracts the shards and sets path to your local root for compatibility with evbench.
This changes the local YAML, so checksum verification is intended for a fresh download,
before extraction. The extraction helper requires Python >=3.11.8 and PyYAML.
The configuration retains channels: 5 and the original split mappings.
In an evbench environment, for example:
python -m evbench.train --data /absolute/path/to/dataset/data.yaml --model yolo11s.pt
License
TODO before public release: document the applicable source-dataset terms and the license for this derived release. No license has been inferred.
- Downloads last month
- -