#Modelzoo

#Classification

network	float	qat	quantization	dataset	input shape	bpu latency (ms)
MobileNetV1	74.12	73.92	73.61	ImageNet	1x3x224x224	0.77
MobileNetV2	72.65	72.51	72.11	ImageNet	1x3x224x224	0.69
ResNet 18	72.04	72.03	72.03	ImageNet	1x3x224x224	1.53
ResNet 50	77.37	76.99	76.94	ImageNet	1x3x224x224	3.06
VargNetV2	73.94	73.56	73.64	ImageNet	1x3x224x224	0.78
EfficientNet-B0	74.31	74.23	74.18	ImageNet	1x3x224x224	0.91
SwinTransformer	80.24	80.15	80.05	ImageNet	1x3x224x224	14.50
MixVarGENet	71.33	71.23	71.04	ImageNet	1x3x224x224	0.56
VargConvert	78.98	78.92	78.89	ImageNet	1x3x224x224	1.51
EfficieNasNetm	80.24	79.99	79.94	ImageNet	1x3x280x280	2.06
EfficieNasNets	76.63	76.23	76.03	ImageNet	1x3x300x300	4.79
ViT-small	79.50	79.40	-	ImageNet	1x3x224x224

Torchvision(Floating point models from the community)：

network	float	qat	quantization	dataset	input shape	bpu latency (ms)
ResNet 18	69.76	69.71	69.73	ImageNet	1x3x224x224	1.59
ResNet 50	76.13	76.07	76.06	ImageNet	1x3x224x224	3.18
MobileNetV2	71.88	71.27	71.27	ImageNet	1x3x224x224	0.85

#Detection

FCOS

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
FCOS-efficientnet	efficientnetb0	36.26	35.79	35.59	MS COCO	1x3x512x512	1.35
FCOS-efficientnet	efficientnetb1	41.37	41.21	40.71	MS COCO	1x3x640x640	2.78
FCOS-efficientnet	efficientnetb2	45.35	45.10	45.00	MS COCO	1x3x768x768	4.32
FCOS-efficientnet	efficientnetb3	48.03	47.65	47.58	MS COCO	1x3x896x896	7.14

DETR

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
DETR-resnet50	resnet50	35.70	31.42	31.31	MS COCO	1x3x800x1333	41.36
DETR-efficientnetb3	efficientnetb3	37.21	35.95	35.99	MS COCO	1x3x800x1333	32.33

FCOS3D

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
FCOS3D-efficientnetb0	efficientnetb0	30.60	30.27	30.31	nuscenes	1x3x512x896	3.51

#Segmentation

UNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
UNet	MobileNetV1	68.02	67.56	67.53	Cityscapes	1x3x1024x2048	2.10

Deeplab

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
Deeplab	EfficientNet-M0	76.30	76.22	76.12	Cityscapes	1x3x1024x2048	4.78
Deeplab	EfficientNet-M1	77.94	77.64	77.65	Cityscapes	1x3x1024x2048	11.23
Deeplab	EfficientNet-M2	78.82	78.65	78.63	Cityscapes	1x3x1024x2048	17.37

FastScnn

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
FastScnn	EfficientNet-B0lite	69.97	69.90	69.88	Cityscapes	1x3x1024x2048	1.92

#OpticalFlow

PwcNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
PwcNet-lg	PwcNet	1.4117	1.4112	1.4075	FlyingChairs	1x6x384x512	12.62

#Lidar

PointPillars

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
PointPillars	SequentialBottleNeck	77.31	76.86	76.76	KITTI3D	150000x4	0.82

CenterPoint

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
CenterPoint	SequentialBottleNeck	58.32	58.11	58.14	nuscenes	1x5x20x40000, 40000x4	23.38

LidarMultiTask

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
LidarMultiTask	MixVarGENet	58.09	57.72	57.62	nuscenes	1x5x20x40000, 40000x4	22.07

#Lane Detection

GaNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
GaNet	MixVarGENet	79.49	78.72	78.72	CuLane	1x3x320x800	1.053

#Multiple Object Track

Motr

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
Motr	efficientnetb3	58.02	57.62	57.76	Mot17	1x3x800x1422, 1x256x2x128, 1x1x1x256, 1x4x2x128	26.40

#Binocular depth estimation

StereoNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
StereoNet	StereoNeck	1.1270	1.1677	1.1685	SceneFlow	1x6x540x960	28.33
StereoNetPlus	MixVarGENet	1.1270	1.1329	1.1351	SceneFlow	2x3x544x960	6.50

#Bev

Bev

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
BevIPM	efficientnetb0	30.59	30.80	30.41	nuscenes det	6x3x512x960, 6x128x128x2	9.76
BevIPM	efficientnetb0	51.47	51.41	50.98	nuscenes seg	6x3x512x960, 6x128x128x2	9.76
BevLSS	efficientnetb0	30.09	30.05	30.01	nuscenes det	6x3x256x704, 10x128x128x2, 10x128x128x2	7.35
BevLSS	efficientnetb0	51.78	51.47	51.46	nuscenes seg	6x3x256x704, 10x128x128x2, 10x128x128x2	7.35
BevGKT	MixVarGENet	28.11	28.12	27.90	nuscenes det	6x3x512x960, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2	23.07
BevGKT	MixVarGENet	48.53	48.02	48.37	nuscenes seg	6x3x512x960, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2	23.07
BevIPM4D	efficientnetb0	37.21	37.19	37.17	nuscenes det	6x3x512x960, 6x128x128x2, 1x64x128x128, 1x128x128x2	10.05
BevIPM4D	efficientnetb0	52.90	53.80	53.77	nuscenes seg	6x3x512x960, 6x128x128x2, 1x64x128x128, 1x128x128x2	10.05
Detr3d	efficientnetb3	34.04	33.87	33.39	nuscenes det	6x3x512x1408	68.54
PETR	efficientnetb3	37.60	37.32	37.31	nuscenes det	6x3x512x1408	228.50
Bevformer-tiny	resnet50	37.00	36.66	-	nuscenes det	6x3x480x800	/
BevCFT	efficientnetb3	32.93	32.68	-	nuscenes det	6x3x512x1408	/

#Keypoint Detection

HeatmapKeypointModel

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
HeatmapKeypointModel	efficientnetb0	94.33	94.30	94.31	carfusion	1x3x128x128	0.83

#Trajectory Prediction

DenseTNT

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)
DenseTNT	vectornet	1.2974	1.2989	1.3038	argoverse 1	30x9x19x32, 30x11x9x64, 30x1x1x96, 30x2x1x2048, 30x1x1x2048	26.45

QCNet

network	backbone	float	qat	quantization	dataset	input shape
qcnet	-	83.84	-	-	argoverse 2	see the list below

The input shape of the QCNet model is:

1x30x50, 1x50x30x30, 1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x80, 1x1x30x80, 1x1x30x80, 1x1x30x10, 1x1x30x10, 1x1x30x10, 1x1x30x10, 1x1x30x30, 1x1x30x30, 1x1x30x30, 1x30x29x128, 1x30x10x128, 1x30x10x128, 1x80, 1x80, 1x1x80x80, 1x1x80x80, 1x1x80x80, 1x1x80x50, 1x1x80x50, 1x1x80x50, 1x80x50, 1x80x50, 1x80x50, 1x80x50, 1x30x30, 1x30x1, 1x1x30x30, 1x1x30x30, 1x1x30x30, 1x1x30x30, 1x1x30x80, 1x1x30x80, 1x1x30x80, 1x1x30x30, 1x1x30x30, 1x1x30x30, 1x80x80.

#Online Map Construction

MapTR

network	backbone	float	qat	quantization	dataset	input shape
MapTRv2	resnet50	0.5859	0.5843	0.5763	nuscenes	6x3x480x800