ilessio-aiflowlab/project_saga

ANIMA Saga — Self-Aware Object Detection via Degradation Manifolds

Paper: arXiv:2602.18394 (Becker et al., 2026)

Implementation by: RobotFlow Labs / AIFLOW Labs

Overview

Saga adds degradation-aware self-awareness to any object detector. A lightweight embedding head trained via multi-layer contrastive learning detects when input quality degrades (blur, noise, rain, fog, compression) — enabling safety-critical systems to flag unreliable perception rather than trusting silent failures.

Results

Metric	Value
AUROC (pristine vs degraded)	0.7991
Detector backbone	yolov10m
Training epochs	7
Embedding dimension	128

Paper Table 1 Reference (YOLOv10-m, COCO mixed degradation)

Severity	1	2	3	4	5
Paper	88.64	89.70	89.75	95.28	97.14
Ours	TBD	TBD	TBD	TBD	TBD

Usage

import torch
from anima_saga.wrappers.detector_registry import PaperDetectorWrapper
from anima_saga.core.prototype import PristinePrototype

# Load model
model = PaperDetectorWrapper("yolov10m", embedding_dim=128, freeze_backbone=True)
model.load_state_dict(torch.load("pytorch/saga_awareness_v1.pth")["model_state_dict"])
model.eval().cuda()

# Load prototype
prototype = PristinePrototype.load("checkpoints/pristine_prototype.pkl")

# Inference
image = torch.randn(1, 3, 640, 640).cuda()  # Your image here
with torch.no_grad():
    embedding = model(image)
    score = prototype.score_cosine(embedding)
    # score ~ 0: pristine, score > 0.5: degraded
    print(f"Degradation score: {score.item():.4f}")

Files

File	Description
pytorch/saga_awareness_v1.pth	PyTorch checkpoint (resume training)
pytorch/saga_awareness_v1.safetensors	SafeTensors (fast loading)
onnx/saga_awareness_v1.onnx	ONNX (cross-platform)
tensorrt/saga_awareness_v1_fp16.trt	TensorRT FP16 (Jetson/L4)
tensorrt/saga_awareness_v1_fp32.trt	TensorRT FP32
checkpoints/pristine_prototype.pkl	Pristine prototype for scoring
configs/paper_cuda.toml	Training config (reproducibility)
logs/training_history.json	Loss curves + metrics

Architecture

Input (640x640) → YOLOv10-m backbone → Multi-layer features
  → 1x1 conv + attention pooling per layer
  → Concatenate → MLP projection → L2 normalize
  → Cosine distance from pristine prototype = degradation score

Citation

@article{becker2026selfaware,
  title={Self-Aware Object Detection via Degradation Manifolds},
  author={Becker, Stefan and Weiss, Simon and H\"ubner, Wolfgang and Arens, Michael},
  journal={arXiv preprint arXiv:2602.18394},
  year={2026}
}

License

Apache 2.0