BlindNav dataset

BlindNav is a private research dataset for detecting Vietnamese street objects that may matter to blind and low-vision pedestrians.

Recommended training release

Use releases/yolo-40-v1/ with khoadd879/BlindNav-model. It is a portable Ultralytics YOLO detection dataset:

releases/yolo-40-v1/
├── dataset.yaml
├── release_manifest.json
├── images/{train,val,test}/
└── labels/{train,val,test}/

Property	Value
Images	1,933
Bounding boxes	11,856
Classes	40
Train	1,546
Validation	193
Test	194
Split seed	42

The 40-class order is copied from the published weights/best.pt checkpoint and validated as an ordered subset of the canonical 49 classes in dataset_config.yaml. Three source JSON files with blocking technical errors were excluded without modifying the source annotations; details are recorded in release_manifest.json.

Nine canonical classes are not part of this checkpoint-compatible release:

hazard.ground_water_hose
hazard.low_wire
hazard.deep_hole
sidewalk_vendor.awning_tarp_canopy
landmark.bus_stop_sign
landmark.atm
infra.public_water_tap
infra.telecom_ground_box
infra.warning_tape

These classes require a future cold-start 49-class model; do not insert them into the 40-class mapping or resume the current checkpoint with a different class order.

Download and validate

hf download khoadd879/BlindNav \
  --repo-type dataset \
  --include "releases/yolo-40-v1/**" \
  --local-dir ./BlindNav-dataset

hf download khoadd879/BlindNav-model \
  weights/best.pt scripts/train_hf_yolo.py \
  --local-dir ./BlindNav-model

python ./BlindNav-model/scripts/train_hf_yolo.py \
  --dataset-dir ./BlindNav-dataset/releases/yolo-40-v1 \
  --model-path ./BlindNav-model/weights/best.pt \
  --validate-only

dataset.yaml intentionally has no absolute path: key. Ultralytics resolves the split directories relative to the YAML file, so the same release works locally, in Google Colab, and in Hugging Face Jobs.

Repository layout

The older root class folders, labels_v2, labels_unified, _mini_yolo, and blindnav_unified are workspace/source or legacy artifacts. They are retained for research provenance but should not be used as a training release.

dataset_config.yaml is the source of truth for the full canonical taxonomy. Class order must not be changed without creating a new dataset/model version and retraining.

Privacy, license, and limitations

The repository is currently private and intended for research. Street images may contain people, vehicle plates, storefronts, or other identifying details. Complete a privacy and data-license review before making the dataset public.

Rare classes are strongly imbalanced; some 40-class entries have only two boxes. This release must not be treated as sufficient evidence of reliable safety-critical performance.