TY-RIST: Tactical YOLO Tricks for Real-time Infrared Small Target Detection
Paper • 2509.22909 • Published
GUNGNIR — TY-RIST: Tactical YOLO for Real-time Infrared Small Target Detection
ANIMA perception module implementing the TY-RIST architecture from arXiv:2509.22909.
Model Details
| Property | Value |
|---|---|
| Architecture | TY-RIST (stride-aware backbone + cascaded CA + NWD loss) |
| Parameters | 2.48M |
| Input | 1x640x640 (single-channel infrared) |
| Detection Heads | P2 (stride 4), P3 (stride 8), P4 (stride 16), P5 (stride 32) |
| Dataset | NUAA-SIRST (427 images) |
| Best Val Loss | 0.4234 |
Training
- Hardware: 2x NVIDIA L4 (23GB) with PyTorch DDP
- Epochs: 200 (11.7 minutes)
- Batch Size: 64/GPU (128 effective)
- Optimizer: AdamW (lr=0.0001, wd=0.05)
- Scheduler: Cosine warmup (5 epochs)
- Precision: FP16 (AMP)
- Pretrained: Random init (YOLOv12n COCO weights only 33.9% compatible)
Available Formats
| Format | File | Size | Description |
|---|---|---|---|
| PyTorch | gungnir_best.pt |
30 MB | Full checkpoint |
| safetensors | gungnir_best.safetensors |
10 MB | Weights only |
| ONNX | gungnir.onnx |
10.6 MB | Opset 18, dynamic batch |
| TensorRT FP16 | gungnir_fp16.engine |
6.7 MB | Built on L4 (Ada) |
| TensorRT FP32 | gungnir_fp32.engine |
14.7 MB | Built on L4 (Ada) |
| MLX | gungnir_mlx.npz |
10 MB | Apple Silicon |
Usage
from anima_gungnir.inference import TorchInferenceEngine
from anima_gungnir.train import load_experiment_config
cfg = load_experiment_config("configs/train/default.yaml")
engine = TorchInferenceEngine(cfg, checkpoint_path="gungnir_best.pt")
detections = engine.predict_image("infrared_frame.png")
Citation
@article{tyrist2025,
title={TY-RIST: Tactical YOLO for Real-time Infrared Small Target Detection},
journal={arXiv preprint arXiv:2509.22909},
year={2025}
}
License
MIT — AIFLOW LABS / RobotFlow Labs