hae-s-ir: open thermal person detector for search-and-rescue drones

A real HIT-UAV test frame (night, 60 m): green — detections with scores, orange — missed people.

hae-s-ir finds people in thermal (LWIR) drone imagery. It has 313 k parameters, the int8 ONNX file is 383 KB, and a 256×320 tile runs in 55 ms on a Raspberry Pi Zero 2 W (4 threads, measured). It was trained from scratch on openly licensed civilian data, and the code and weights are MIT-licensed.

Live demo (runs in your browser): https://huggingface.co/spaces/alexshynkarenk0/hae-demo
Related model: hae-n-ir-ext-kdf (code: https://github.com/sashapiscivca-svg/hae-n-ir-ext-kdf), a real-time model for forest and wilderness footage
Code, documentation and training recipe: https://github.com/sashapiscivca-svg/hae
Full model card (EN): https://github.com/sashapiscivca-svg/hae/blob/main/MODEL_CARD.md · (UK): https://github.com/sashapiscivca-svg/hae/blob/main/MODEL_CARD.uk.md
Responsible use: https://github.com/sashapiscivca-svg/hae/blob/main/RESPONSIBLE_USE.md

Decision support only. The model misses people and raises false alarms. Never use it as the only means of search, and never call off a search because it found nothing.

Files

File	Purpose
`hae-s-ir_lwir_int8.onnx`	deployment (int8, fp32 head), 383 KB
`hae-s-ir_lwir.onnx`	fp32 reference, 1.2 MB
`hae-s-ir.safetensors`	PyTorch weights for fine-tuning with the `hae` package
`config.json`	architecture and full training recipe
`metrics.json`	all evaluation numbers, including threshold sweeps

Usage

pip install "hae[hub] @ git+https://github.com/sashapiscivca-svg/hae"
hae predict hae-s-ir frames/            # downloads this repository's int8 model on first use

from hae.detector import Detector
det = Detector("hae-s-ir")              # preprocessing, ×1.5 scale and threshold from the ONNX metadata
res = det.predict("thermal_frame.png")  # 8-bit grayscale, white-hot
print(res.boxes, res.conf)

Without the hae package: the input t is float32 (B,1,H,W) with H and W multiples of 32. Normalise each frame as (t − mean) / max(std, 1) on the 0..255 scale, then upscale ×1.5. The outputs are cls (B,1,H/4,W/4) logits and box (B,4,H/4,W/4). Decode by taking 3×3 local maxima of sigmoid(cls); each box is softplus(box) × 4 around the cell centre ((j+0.5)·4, (i+0.5)·4). Then divide by the scale.

Results (int8, IoU ≥ 0.5)

Set	AP50	Recall / precision / false alarms per frame at the balanced preset (0.50)
HIT-UAV test (579 frames, 2,611 people)	0.757	0.73 / 0.75 / 1.09
AIResQ val (502 frames, 933 people, ~8 px tall)	0.230	0.52 / 0.24 / 3.11
WiSARD-IR val (transfer; 1,633 frames, 2,439 people)	0.087	0.17 / 0.16 / 1.28

The weak transfer to WiSARD-IR (other cameras and sites) is the main limitation. Validate on your own camera and fine-tune if possible.

Training data and licence

The model was trained on HIT-UAV trainval (CC0 1.0) and AIResQ Benchmark train (CC BY 4.0), and evaluated additionally on WiSARD thermal (MIT). Attribution: Suo et al., Scientific Data 10, 227 (2023); Büttner, AIResQ Benchmark, Zenodo doi:10.5281/zenodo.17405074 (CC BY 4.0); Broyles, Hayner, Leung, IROS 2022. The model weights are released under the MIT License, © 2026 The hae authors.

Citation