Enlight NPU :: Model Zoo

Model Name	Type	Input Size	Top1 Acc (FP32)	Top5 Acc (FP32)	Top1 Acc (INT8)	Top5 Acc (INT8)	Giga MAC	Preview	Source
mobilenet_v2	TFLITE	224x224	71.090	90.272	69.738	89.482	0.3		google
mobilenet_v2	ONNX	224x224	71.190	90.206	68.966	88.676	0.3		torchvision
mobilenet_v2	ONNX	224x224	70.958	89.938	69.422	88.912	0.4		onnx
mobilenet_v2_1.4	TFLITE	224x224	74.184	91.920	74.180	91.922	0.6		google
resnet18	ONNX	224x224	68.918	88.826	67.002	87.628	1.8		torchvision
resnet34	ONNX	224x224	72.770	90.966	71.020	90.024	3.7		torchvision
resnet50	ONNX	224x224	75.590	92.784	74.216	91.922	4.1		torchvision
efficientnet-lite-b0	TFLITE	224x224	73.558	91.578	71.606	90.522	1.7		google

Model Name	Type	Input Size	mAP (FP32)	mAP (INT8)	Giga MAC	Preview	Source
ssd_mobilenet_v2	ONNX	300x300	70.7	70.1	0.7	preview	in-house
		320x320			0.8
		512x512	74.8	74.0	1.9
YOLOv2-voc*	DarkNet	416x416	75.5	74.8	14.7		in-house
YOLOv2-tiny-voc	DarkNet	416x416	55.6	55.4	3.5		cfg, weights

* YOLOv2 is trained after replacing buggy reorg layer to reorg3d layer.

Model Name	Type	Input Size	AP (FP32)	AP (INT8)	Giga MAC	Preview	Source
ssd_mobilenet_v2_320x320	TFLITE	320x320	33.6@IoU=0.5 19.2@IoU=0.5:0.95				link
YOLOv2*	DarkNet	416x416	54.8@IoU=0.5 28.2@IoU=0.5:0.95	54.2@IoU=0.5 27.5@IoU=0.5:0.95	31.5		in-house
YOLOv2-tiny	DarkNet	416x416	25.9@IoU=0.5 10.0@IoU=0.5:0.95	25.5@IoU=0.5 9.7@IoU=0.5:0.95	2.7		cfg, weights
YOLOv3	DarkNet	320x320	62.7@IoU=0.5 34.6@IoU=0.5:0.95	62.9@IoU=0.5 34.6@IoU=0.5:0.95	19.5		cfg, weights
		416x416	67.1@IoU=0.5 37.6@IoU=0.5:0.95	67.7@IoU=0.5 37.9@IoU=0.5:0.95	32.9
		608x608	68.0@IoU=0.5 38.3@IoU=0.5:0.95	68.9@IoU=0.5 38.7@IoU=0.5:0.95	70.3
YOLOv3-tiny	DarkNet	416x416	20.0@IoU=0.5 9.3@IoU=0.5:0.95	19.7@IoU=0.5 9.1@IoU=0.5:0.95	2.5		cfg, weights
YOLOv3-spp	DarkNet	608x608	69.9@IoU=0.5 42.0@IoU=0.5:0.95	69.3@IoU=0.5 41.3@IoU=0.5:0.95	70.7		cfg, weights
YOLOv4	DarkNet	320x320	62.5@IoU=0.5 39.7@IoU=0.5:0.95	62.2@IoU=0.5 39.0@IoU=0.5:0.95	17.8		cfg, weights
		416x416	70.5@IoU=0.5 46.1@IoU=0.5:0.95	70.2@IoU=0.5 45.3@IoU=0.5:0.95	30.1
		512x512	73.6@IoU=0.5 48.8@IoU=0.5:0.95	72.9@IoU=0.5 47.8@IoU=0.5:0.95	45.5
		608x608	74.6@IoU=0.5 49.8@IoU=0.5:0.95	74.3@IoU=0.5 48.5@IoU=0.5:0.95	64.2
YOLOv4-csp	DarkNet	512x512	64.2@IoU=0.5 45.2@IoU=0.5:0.95	63.8@IoU=0.5 43.6@IoU=0.5:0.95	38.5		cfg, weights
		608x608	65.4@IoU=0.5 46.2@IoU=0.5:0.95	65.0@IoU=0.5 44.5@IoU=0.5:0.95	45.7
YOLOv4-tiny	DarkNet	416x416	41.4@IoU=0.5 21.4@IoU=0.5:0.95	40.7@IoU=0.5 20.7@IoU=0.5:0.95	3.5		cfg, weights

Model Name	Type	Input Size	mIoU / Pixel Acc / Mean Acc (FP32)	mIoU / Pixel Acc / Mean Acc (INT8)	Giga MAC	Preview	Source
PIDNet-S	ONNX	2048x1024	78.1 / 96.1 / 85.8	77.9 / 96.0 / 85.6	50.5		link

Model Name	Type	Input Size	mask mAP* (FP32)	mask mAP* (INT8)	Giga MAC	Preview	Source
Yolact-MobileNetV2	ONNX	512x512	36.40@IoU=0.5 21.11@IoU=0.5:0.95	35.67@IoU=0.5 20.71@IoU=0.5:0.95	29.6		link
		640x640	38.23@IoU=0.5 22.46@IoU=0.5:0.95	37.39@IoU=0.5 21.97@IoU=0.5:0.95	46.1

Model Name	Type	Input Size	Giga MAC	Preview	Source
Blaze Face	TFLITE*	128x128	0.046		link

Model Name	Type	Input Size	OKS (FP32)	OKS (INT8)	Giga MAC	Preview	Source
YOLO-pose	ONNX	640x640	64.5	62.7	10.1		link

Model Name	Type	Input Size	F1-Score (FP32)	F1-Score (INT8)	Giga MAC	Preview	Source
UFLD v1	ONNX	800x288	69.1	68.6	8.95		link

Model Name	Type	Input Size	ARE (FP32)	ARE (INT8)	Giga MAC	Preview	Source
Monodepth2	ONNX	640x192	11.1	12.6	8.74		link

• Classification Networks
  • MobileNetV2: Inverted Residuals and Linear Bottlenecks
  • Deep Residual Learning for Image Recognition
• Object Detection Networks
  • SSD: Single Shot MultiBox Detector
  • YOLO9000: Better, Faster, Stronger
  • YOLOv3: An Incremental Improvement
  • YOLOv4: Optimal Speed and Accuracy of Object Detection
  • YOLOX: Exceeding YOLO Series in 2021
  • YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
• Semantic Segmentation Networks
  • PIDNet: A Real-time Semantic Segmentation Network Inspired by PID Controllers
• Instance Segmentation Networks
  • YOLACT: Real-time Instance Segementation
  • YOLACT++: Better Real-time Instance Segementation
  • YolactEdge: Real-time Instance Segmentation on the Edge
• Face Detection Networks
  • BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs
• Pose Estimation Networks
  • YOLO-Pose: Enhancing YOLO for Multi Person Estimation Using Object Keypoint Similarity Loss
• Lane Detection Networks
  • Ultra Fast Structure-aware Deep Lane Detection
• Monocular Depth Estimation Networks
  • Digging Into Self-Supervised Monocular Depth Estimation
• DataSets
  • ImageNet is an image database organized according to the WordNet hierarchy (currently only the nouns), in which each node of the hierarchy is depicted by hundreds and thousands of images.
  • PASCAL Visual Object Classes (VOC)
  • COCO is a large-scale object detection, segmentation, and captioning dataset.
  • Cityscape: Benchmark suite and evaluation server for pixel-level, instance-level, and panoptic semantic labeling
  • CULane: CULane is a large scale challenging dataset for academic research on traffic lane detection.
  • KITTI: Dataset for stereo, optical flow, visual odometry, 3D object detection and 3D tracking
• Metrics
  • mAP: The PASCAL Visual Object Classes (VOC) Challenge
  • Mask mAP: Mask Scoring R-CNN
  • Average Precision (AP): Detection Evaluation (COCO)
  • Object Keypoint Similarity (OKS): Keypoint Evaluation (COCO)
  • F1-Score: F-score is a measure of a test's accuracy
  • Absolute Relative Error (ARE)