I’ve been obsessing over edge-ready anomaly detection—tiny footprint, big accuracy, no GPU. So I pitted my lean PaDiM pipeline against Anomalib on MVTec (bottle), CPU-only. The results made me smile 😎
✨ Headline Results
⚡ Latency: 22.5 ms/image (44.5 FPS) vs 105.3 ms (9.5 FPS)
🎯 Image AUROC: 0.9968 (mine) vs 0.9960
🧩 Pixel AUROC: 0.9837 (mine) vs 0.9869
💾 Model+stats size: 15.25 MB vs 40.47 MB
TL;DR: ~4–5× faster on CPU with matching image-level accuracy and a much smaller footprint. Anomalib edges me on pixel AUROC by a whisker (I’m coming for it 😉).
Deep knowledge
I’ve been obsessing over edge-ready anomaly detection—tiny footprint, big accuracy, no GPU. So I pitted my lean PaDiM pipeline against Anomalib on MVTec (bottle), CPU-only. The results made me smile 😎
Checkout: github.com/DeepKnowledge1/AnomaVision
✨ Headline Results
⚡ Latency: 22.5 ms/image (44.5 FPS) vs 105.3 ms (9.5 FPS)
🎯 Image AUROC: 0.9968 (mine) vs 0.9960
🧩 Pixel AUROC: 0.9837 (mine) vs 0.9869
💾 Model+stats size: 15.25 MB vs 40.47 MB
TL;DR: ~4–5× faster on CPU with matching image-level accuracy and a much smaller footprint. Anomalib edges me on pixel AUROC by a whisker (I’m coming for it 😉).
#AnomalyDetection #ComputerVision #EdgeAI #MLOps #PyTorch #Anomalib #PaDiM #ONNX #TensorRT #AIEngineering #ModelOptimization #DeployOnCPU #GPU
1 week ago | [YT] | 2