VoxHOD-NC: drone person detector with extra Turkish footage (non-commercial)

VoxHOD-NC is the non-commercial branch of VoxHOD, a tiny int8 person detector for drone search and rescue on ESP32-S3-class microcontrollers.

Türkçe: VoxHOD-NC, VoxHOD'un ticari olmayan koludur. Ferhat Tatık'ın izniyle onun drone görüntüleri eklenerek eğitildi; bu yüzden yalnızca ticari olmayan kullanım içindir (CC BY-NC 4.0). Ticari kullanım gerekiyorsa, bu görüntüleri hiç görmemiş VoxHOD modelini kullanın (CC BY 4.0). Güncel sürüm v2: Ferhat Tatık'ın test videolarında insan bulma oranı v3'e göre %32,9'dan %62,2'ye çıktı, HERIDAL ve WiSARD'da da v3'ün önünde. Yanlış alarm o görüntülerde hâlâ çok; deneysel bir sürümdür.

Which one do I need?

	VoxHOD	VoxHOD-NC (this repo)
Training data	HERIDAL only	HERIDAL plus drone footage by Ferhat Tatık
Model licence	CC BY 4.0, commercial use allowed	CC BY-NC 4.0, non-commercial only
Status	v3 (197 KB) and nano (75 KB)	v2, experimental (197 KB int8, same architecture as v3)

The footage is used with the photographer's written permission for non-commercial use. The footage itself is not published here or anywhere else, and its rights stay with Ferhat Tatık.

v2 (2026-11-06)

How it was made:

1. Frames were taken every 2 s from 14 of Ferhat Tatık's drone videos (plateaus, coast, mountains, an enduro race, a marathon, earthquake rescue footage).
2. The videos were split by video: 3 held-out test videos (an earthquake scene, a lake valley, a city marathon) are never trained on.
3. People were boxed automatically by a large COCO person detector (teacher: torchvision Faster R-CNN ResNet50-FPN v2) on every 3rd frame. There was no manual review. All 11 training videos are labelled now (1,706 frames, 6,866 training tiles; v0 used 4 videos and 2,204 tiles). Confident v3 hits far from any teacher person are added as hard negatives.
4. Training: v1 continued from VoxHOD v3 on HERIDAL + the new tiles (lr 5e-4, ~170 min float, 45 min quantization-aware training). v2 continued from v1 (lr 3e-4, ~120 min float, 30 min quantization-aware training).

Results, int8 QAT model, threshold 0.2 (the same fixed threshold as VoxHOD). Recall / precision in %, best F1 over all thresholds in brackets:

Test set	v3	NC v0	NC v2
HERIDAL test	83.9 / 20.4 (0.569)	82.0 / 21.0 (0.579)	85.2 / 19.4 (0.603)
WiSARD clip (external)	61.3 / 19.1 (0.449)	60.5 / 20.8 (0.469)	70.5 / 22.4 (0.503)
Ferhat test videos, 320 px crops	32.9 / 5.4	52.2 / 6.4 (0.162)	62.2 / 5.7 (0.166)
Ferhat test videos, 640 px crops	14.9 / 5.9	27.3 / 9.0 (0.146)	35.6 / 8.4 (0.145)

v1 (not published) found more people than v0 but raised more false alarms (best F1 0.145 and 0.132 on the two Ferhat sets). v2 keeps v1's recall and recovers most of the precision.

What this does and does not show:

• On the new kind of footage, recall rises by about 29 points (320 px) and 21 points (640 px) over v3. On HERIDAL it is about 1 point higher, and on the WiSARD clip, which neither model trained on, about 9 points higher.
• Precision on the new footage is still very low (6-8%), with several false alarms per tile. These are mostly oblique, cinematic views with crowds, boats and buildings, which are far from the top-down search-and-rescue images VoxHOD was built for.
• The Ferhat test labels come from the teacher, not from people. The teacher misses some people (swimmers, people under trees), so some "false alarms" may be real people, and recall is measured against the teacher's view.
• Tiles whose people are outside 3-64 px on the 160 px tile were left out of the Ferhat test, since VoxHOD is not built for that scale.

Next: check a sample of teacher boxes by hand, cut the false alarms on oblique footage, and keep the test videos untouched.

Files

File	What
weights/voxhod_nc_qat_int8.tflite	int8, quantization-aware trained (use this one)
weights/voxhod_nc_int8.tflite	int8, post-training quantization
weights/voxhod_nc.keras	float Keras model
weights/evaluation_nc.json	all thresholds, all test sets
weights/evaluation_v3_ferhat_test.json	v3 on the Ferhat test videos, for comparison
scripts/	teacher labelling, v3 hard negatives and tile building

Input and output are the same as VoxHOD v3: a 160x160 RGB tile in, a 40x40 person-centre heatmap plus box size out.

Code

Training and evaluation code is shared with VoxHOD and lives in ErdemWilkinson/VoxHOD (MIT). The scripts here only build the extra data; they need the footage, which is not published.

Thanks

Thanks to Ferhat Tatık (@ferhattatik_dronoloji, YouTube @dronoloji) for the drone footage. @ferhattatik_dronoloji teşekkürler!

Licence

Model weights: CC BY-NC 4.0. See MODEL_LICENSE.md. Scripts: MIT, as in VoxHOD.

VoxHOD-NC is built for search and rescue, to help find people who need help. It is not intended for targeting or harming people.