USV-DETR

Detection of rodent ultrasonic vocalizations in spectrograms. Built on RT-DETR with a high-resolution P2 feature layer and the DEIM training framework.

Paper: USV-DETR: High-Resolution and Densely Supervised Detection of Ultrasonic Vocalizations (Interspeech 2026)
Code: https://github.com/weiyilan9/USV-DETR

Files

USV-DETR.pth: training checkpoint, usable for both inference and fine-tuning
USV-DETR.yml: model config, requires the RT-DETRv4 codebase
config.json: metadata only

Training data

SqueakOut: A public mouse USV dataset. Five strains (C57Bl6/J, NZO/HlLtJ, 129S1/SvImJ, NOD/ShiLtJ, PWK/PhJ), postnatal day 5 to 15.

Input format

640 x 640 RGB spectrogram image
Pixel values divided by 255, no mean/std normalization
Plain stretch resize, not letterbox
num_classes = 1

Performance

AP 78.7, AP50 94.5, AP75 86.1, APs 74.9 (COCO metrics) on the SqueakOut test split.

Usage

from huggingface_hub import hf_hub_download

hf_hub_download("yilanwei/USV-DETR", "config.json")   # metadata
ckpt = hf_hub_download("yilanwei/USV-DETR", "USV-DETR.pth")
cfg  = hf_hub_download("yilanwei/USV-DETR", "USV-DETR.yml")

Inference reads checkpoint["ema"]["module"]. See the GitHub repo for the full pipeline.

Limitations

The model reads spectrograms, not audio, so different STFT settings, frequency ranges, or colormaps may fall outside the training distribution. Performance on other species, age ranges, or recording setups has not been validated.

Downloads last month: -