10.5. Modelzoo

10.5.1. Classification

network	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
MobileNetV1	74.12	73.92	73.61	ImageNet	1x3x224x224	0.983	1318.82
MobileNetV2	72.65	72.51	72.11	ImageNet	1x3x224x224	0.888	1508.56
ResNet 18	72.04	72.03	72.03	ImageNet	1x3x224x224	2.507	443.53
ResNet 50	77.37	76.99	76.94	ImageNet	1x3x224x224	4.917	214.31
VargNetV2	73.94	73.56	73.64	ImageNet	1x3x224x224	0.871	1557.39
EfficientNet-B0	74.31	74.23	74.18	ImageNet	1x3x224x224	1.183	1047.60
SwinTransformer	80.24	80.15	80.05	ImageNet	1x3x224x224	22.76	44.59
MixVarGENet	71.33	71.23	71.04	ImageNet	1x3x224x224	1.023	1282.44
VargConvert	78.98	78.92	78.89	ImageNet	1x3x224x224	3.409	316.88
EfficieNasNetm	80.24	79.99	79.94	ImageNet	1x3x300x300	3.531	305.12
EfficieNasNets	76.63	76.23	76.03	ImageNet	1x3x280x280	1.536	766.52

Torchvision(浮点模型来自社区)：

network	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
ResNet 18	69.76	69.71	69.73	ImageNet	1x3x224x224	2.479	403.38
ResNet 50	76.13	76.07	76.06	ImageNet	1x3x224x224	4.989	200.44
MobileNetV2	71.88	71.27	71.27	ImageNet	1x3x224x224	1.32	1249.79

10.5.2. Detection

RetinaNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
Retinanet-vargnetv2	vargnetv2	31.51	31.21	31.20	MS COCO	1x3x1024x1024	99.92	10.05

YOLOv3

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
YOLOv3-MobileNetv1	mobilenetv1	76.57	75.62	75.61	VOC	1x3x416x416	10.83	96.46
YOLOv3-VarGDarknet	VarGDarknet	33.90	33.60	33.36	COCO	1x3x416x416	19.03	52.54

FCOS

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
FCOS-efficientnet	efficientnetb0	36.26	35.79	35.59	MS COCO	1x3x512x512	3.164	345.95
FCOS-efficientnet	efficientnetb1	41.37	41.21	40.71	MS COCO	1x3x640x640	7.909	131.95
FCOS-efficientnet	efficientnetb2	45.35	45.10	45.00	MS COCO	1x3x768x768	13.91	73.76
FCOS-efficientnet	efficientnetb3	48.03	47.65	47.58	MS COCO	1x3x896x896	25.06	38.92

DETR

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
DETR-resnet50	resnet50	35.70	31.42	31.31	MS COCO	1x3x800x1333	128.4	7.81
DETR-efficientnetb3	efficientnetb3	37.21	35.95	35.99	MS COCO	1x3x800x1333	91.82	10.93

FCOS3D

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
FCOS3D-efficientnetb0	efficientnetb0	30.62	30.27	30.38	nuscenes	1x3x512x896	11.26	92.79

10.5.3. Segmentation

UNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
UNet	MobileNetV1	68.02	67.56	67.53	Cityscapes	1x3x1024x2048	5.775	181.81

Deeplab

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS (dual core)
Deeplab	EfficientNet-M0	76.30	76.22	76.12	Cityscapes	1x3x1024x2048	21.25	47.73
Deeplab	EfficientNet-M1	77.94	77.64	77.65	Cityscapes	1x3x1024x2048	44.36	22.69
Deeplab	EfficientNet-M2	78.82	78.65	78.63	Cityscapes	1x3x1024x2048	70.67	14.20

FastScnn

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
FastScnn	EfficientNet-B0lite	69.97	69.90	69.88	Cityscapes	1x3x1024x2048	8.001	129.36

10.5.4. OpticalFlow

PwcNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
PwcNet-lg	PwcNet	1.4117	1.4112	1.4075	FlyingChairs	1x6x384x512	41.04	24.61

10.5.5. Lidar

PointPillars

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
PointPillars	SequentialBottleNeck	77.31	76.86	76.76	KITTI3D	150000x4	55.88	25.66

CenterPoint

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
CenterPoint	SequentialBottleNeck	58.32	58.11	58.14	nuscenes	1x5x20x40000, 40000x4	63.26	18.99

LidarMultiTask

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
LidarMultiTask	MixVarGENet	58.09	57.72	57.62	nuscenes	1x5x20x40000, 40000x4	44.85	27.77

注解

PointPillars 的指标是 Box3d Moderate 这项。

10.5.6. Lane Detection

GaNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
GaNet	MixVarGENet	79.49	78.72	78.72	CuLane	1x3x320x800	3.741	291.64

10.5.7. Multiple Object Track

Motr

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
Motr	efficientnetb3	58.02	57.62	57.76	Mot17	1x3x800x1422, 1x256x2x128, 1x1x1x256, 1x4x2x128	75.75	12.26

10.5.8. Binocular depth estimation

StereoNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
StereoNet	StereoNeck	1.1270	1.1677	1.1685	SceneFlow	1x6x540x960	90.44	11.13
StereoNetPlus	MixVarGENet	1.1270	1.1329	1.1351	SceneFlow	2x3x544x960	18.6	55.23

10.5.9. Bev

BevIPM

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
BevIPM	efficientnetb0	30.59	30.80	30.41	nuscenes det	6x3x512x960, 6x128x128x2	40.34	25.45
BevIPM	efficientnetb0	51.47	51.41	50.98	nuscenes seg	6x3x512x960, 6x128x128x2	40.34	25.45
BevLSS	efficientnetb0	30.09	30.05	30.01	nuscenes det	6x3x256x704, 10x128x128x2, 10x128x128x2	32.35	31.91
BevLSS	efficientnetb0	51.78	51.47	51.46	nuscenes seg	6x3x256x704, 10x128x128x2, 10x128x128x2	32.35	31.91
BevGKT	MixVarGENet	28.11	28.12	27.90	nuscenes det	6x3x512x960, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2	93.56	10.86
BevGKT	MixVarGENet	48.53	48.02	48.37	nuscenes seg	6x3x512x960, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2	93.56	10.86
BevIPM4D	efficientnetb0	37.24	37.19	37.31	nuscenes det	6x3x512x960, 6x128x128x2, 1x64x128x128, 1x128x128x2	43.51	23.49
BevIPM4D	efficientnetb0	52.90	53.80	53.86	nuscenes seg	6x3x512x960, 6x128x128x2, 1x64x128x128, 1x128x128x2	43.51	23.49
Detr3d	efficientnetb3	34.04	33.87	33.39	nuscenes det	6x3x512x1408, 6x2x4x256, 6x2x4x256, 6x2x4x256, 6x2x4x256, 1x24x4x256	196.9	5.078
BevCFT	efficientnetb3	32.93	32.68	32.63	nuscenes det	6x3x512x1408	179.9	5.558

10.5.10. Keypoint Detection

HeatmapKeypointModel

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
HeatmapKeypointModel	efficientnetb0	94.33	94.30	94.31	carfusion	1x3x128x128	0.958	1377.6

10.5.11. Trajectory Prediction

DenseTNT

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
DenseTNT	vectornet	1.2974	1.2989	1.3038	argoverse 1	30x9x19x32, 30x11x9x64, 30x1x1x96, 30x2x1x2048, 30x1x1x2048	52.77	20.42