AA-YOLO introduces an Anomaly-Aware Detection Head (AADH) that reformulates IR target detection as a statistical anomaly detection problem. The AADH adds only +0.2M parameters and +5% FLOPs to any YOLO backbone. Key components:

Spatial Filtering Block -- Two Conv1d layers (3x3) with BatchNorm/ReLU, compressing to C=8 channels
Statistical Anomaly Test -- Multivariate hypothesis test computing -ln(F_mu2) anomaly scores
Scaled Sigmoid Activation -- Temperature-scaled sigmoid (alpha=0.001) for sharp anomaly response
Backbone-agnostic -- Drop-in head for YOLOv5/v7/v9 backbones

This checkpoint uses the AA-YOLOv7t backbone (YOLOv8n architecture as v7-tiny proxy).

Training Summary

Parameter	Value
Dataset	SIRST (427 images: 256 train / 85 val / 86 test)
Backbone	AA-YOLOv7t (YOLOv8n + AADH)
Epochs	100
Optimizer	SGD (lr=0.01, momentum=0.937, weight_decay=0.0005)
Scheduler	Cosine annealing, 3 epoch warmup
Precision	FP16 mixed precision
Best val_loss	1.0874 (epoch 73)
Final train_loss	1.0715
Backbone freeze	First 10 epochs
Hardware	NVIDIA L4 (23GB)
Training time	~2 min (100 epochs)

Loss Curve

Epoch 0: train=2.60, val=2.50
Epoch 25: train=1.14, val=1.15
Epoch 50: train=1.09, val=1.10
Epoch 73 (best): train=1.08, val=1.09
Epoch 99: train=1.07, val=1.09

Exported Formats

Format	File	Size	Use Case
PyTorch (.pth)	`pytorch/fenris_v0.1.0.pth`	12 MB	Training, fine-tuning
SafeTensors	`pytorch/fenris_v0.1.0.safetensors`	12 MB	Fast loading, safe
ONNX (opset 18)	`onnx/fenris_v0.1.0_allinone.onnx`	12.3 MB	Cross-platform inference
TensorRT FP32	`tensorrt/fenris_v0.1.0_fp32.engine`	~17 MB	Full precision inference
TensorRT FP16	`tensorrt/fenris_v0.1.0_fp16.engine`	~8 MB	Edge deployment (Jetson/L4)
Training checkpoint	`checkpoints/best.pth`	24 MB	Resume training (includes optimizer state)

Quick Start

python

import torch
from safetensors.torch import load_file

# Load SafeTensors (recommended)
state_dict = load_file("pytorch/fenris_v0.1.0.safetensors")

# Or load PyTorch
state_dict = torch.load("pytorch/fenris_v0.1.0.pth", map_location="cpu")

ONNX Inference

python

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("onnx/fenris_v0.1.0_allinone.onnx")
input_data = np.random.randn(1, 3, 256, 256).astype(np.float32)
outputs = session.run(None, {"ir_input": input_data})

TensorRT Inference

python

import tensorrt as trt

runtime = trt.Runtime(trt.Logger(trt.Logger.WARNING))
with open("tensorrt/fenris_v0.1.0_fp16.engine", "rb") as f:
    engine = runtime.deserialize_cuda_engine(f.read())

Datasets

SIRST (427 images, 256x256) -- Primary benchmark
IRSTD-1k (1,000 images, 512x512) -- Extended evaluation
VEDAI (1,200 images, IR satellite) -- Cross-domain evaluation

Target Metrics (Paper)

Metric	Target	Dataset
F1	>= 98.1%	SIRST
AP_s	>= 79.8%	SIRST
AP	>= 72.0%	IRSTD-1k
AP	>= 65.0%	VEDAI

ANIMA Module

Module name: fenris
Category: perception.foundation
Container: ghcr.io/robotflow-labs/anima-fenris:0.1.0
API port: 8080
Health: /health, /ready, /info

License

Apache 2.0 -- Robot Flow Labs / AIFLOW LABS LIMITED

ilessio-aiflowlab/project_fenris

FENRIS -- AA-YOLO: Anomaly-Aware YOLO for Infrared Small Target Detection

Paper

Architecture

Training Summary

Loss Curve

Exported Formats

Quick Start

ONNX Inference

TensorRT Inference

Datasets

Target Metrics (Paper)

ANIMA Module

License

ilessio-aiflowlab/
project_fenris