模型仓库modelzoo

Classification

network	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
mobilenetv1_imagenet	74.12	73.92	73.61	ImageNet	1x3x224x224	0.42	5069.83
mobilenetv2_imagenet	72.65	72.51	72.11	ImageNet	1x3x224x224	0.45	4541.88
resnet18_imagenet	72.04	72.03	72.03	ImageNet	1x3x224x224	0.65	2346.22
resnet50_imagenet	77.37	76.99	76.94	ImageNet	1x3x224x224	1.12	1117.93
vargnetv2_imagenet	73.94	73.56	73.64	ImageNet	1x3x224x224	0.50	3763.68
efficientnet_imagenet	74.31	74.23	74.18	ImageNet	1x3x224x224	0.49	3887.20
horizon_swin_transformer_imagenet	80.24	80.15	80.05	ImageNet	1x3x224x224	4.19	257.265367
mixvargenet_imagenet	71.33	71.23	71.04	ImageNet	1x3x224x224	0.44	4728.13
efficientnasnetm_imagenet	80.24	79.99	79.94	ImageNet	1x3x280x280	0.99	1305.20
efficientnasnets_imagenet	76.63	76.23	76.03	ImageNet	1x3x300x300	0.56	2943.99
vit_small_imagenet	79.50	79.40	77.86	ImageNet	1x3x224x224	2.35	472.41
henet_tinye_imagenet	77.68	77.22	76.92	ImageNet	1x3x224x224	0.60	2713.99
henet_tinym_imagenet	78.38	77.95	77.62	ImageNet	1x3x224x224	0.48(J6M)	3451.43(J6M)

Detection

FCOS

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
fcos_efficientnetb0_mscoco	efficientnetb0	36.26	35.79	35.59	MS COCO	1x3x512x512	-	-
fcos_efficientnetb1_mscoco	efficientnetb1	41.37	41.21	40.71	MS COCO	1x3x640x640	2.44	487.89
fcos_efficientnetb2_mscoco	efficientnetb2	45.35	45.10	45.00	MS COCO	1x3x768x768	3.46	325.84
fcos_efficientnetb3_mscoco	efficientnetb3	48.03	47.65	47.58	MS COCO	1x3x896x896	5.75	187.30

DETR

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
detr_resnet50_mscoco	resnet50	35.70	31.42	31.31	MS COCO	1x3x800x1333	28.63	35.27
detr_efficientnetb3_mscoco	efficientnetb3	37.21	35.95	35.99	MS COCO	1x3x800x1333	22.20	45.64

Deform DETR

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
deform_detr_resnet50_mscoco	resnet50	44.34	44.65	44.80	MS COCO	1x3x800x1333	222.42	4.51

FCOS3D

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
fcos3d_efficientnetb0_nuscenes	efficientnetb0	30.60	30.27	30.31	nuscenes	1x3x512x896	3.03	409.89

Segmentation

UNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
unet_mobilenetv1_cityscapes	MobileNetV1	68.02	67.56	67.53	Cityscapes	1x3x1024x2048	1.64	783.06

Deeplab

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
deeplabv3plus_efficientnetm0_cityscapes	EfficientNet-M0	76.30	76.22	76.12	Cityscapes	1x3x1024x2048	4.83	218.18
deeplabv3plus_efficientnetm1_cityscapes	EfficientNet-M1	77.94	77.64	77.65	Cityscapes	1x3x1024x2048	9.21	111.76
deeplabv3plus_efficientnetm2_cityscapes	EfficientNet-M2	78.82	78.65	78.63	Cityscapes	1x3x1024x2048	13.77	74.04

FastScnn

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
fastscnn_efficientnetb0tiny_cityscapes	EfficientNet-B0lite	69.97	69.90	69.88	Cityscapes	1x3x1024x2048	2.27	493.85

Lidar

PointPillars

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
pointpillars_kitti_car	SequentialBottleNeck	77.31	76.86	76.76	KITTI3D	150000x4	185.42	28.12

CenterPoint

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
centerpoint_pointpillar_nuscenes	SequentialBottleNeck	58.32(NDS) 48.04(MAP)	58.11(NDS) 47.85(MAP)	58.14(NDS) 47.81(MAP)	nuscenes det	1x5x20x40000, 40000x4	55.26	100.62

LidarMultiTask

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
centerpoint_mixvargnet_multitask_nuscenes	MixVarGENet	58.09(NDS) 47.27(MAP) 91.28(MeanIOU)	57.72(NDS) 46.76(MAP) 91.18(MeanIOU)	57.53(NDS) 46.28(MAP) 91.21(MeanIOU)	nuscenes det && seg	1x5x20x40000, 40000x4	53.56	125.52

注解

PointPillars 的指标是 Box3d Moderate 这项。

Lane Detection

GaNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
ganet_mixvargenet_culane	MixVarGENet	79.49	78.72	78.72	CuLane	1x3x320x800	0.94	1445.91

Multiple Object Track

Motr

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
motr_efficientnetb3_mot17	efficientnetb3	58.02	57.62	57.76	Mot17	1x3x800x1422, 1x256x2x128, 1x1x1x256, 1x4x2x128	14.99	68.31

Binocular depth estimation

StereoNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
stereonetplus_mixvargenet_sceneflow	MixVarGENet	1.1270	1.1329	1.1351	SceneFlow	2x3x544x960	5.12	208.22

Bev

Bev

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
bev_ipm_efficientnetb0_multitask_nuscenes	efficientnetb0	30.54(NDS) 21.70(MAP) 51.45(MeanIOU)	30.80(NDS) 21.66(MAP) 51.47(MeanIOU)	30.26(NDS) 21.56(MAP) 50.99(MeanIOU)	nuscenes det && seg	6x3x512x960, 6x128x128x2	9.52	112.59
bev_lss_efficientnetb0_multitask_nuscenes	efficientnetb0	30.06(NDS) 20.62(MAP) 51.80(MeanIOU)	30.10(NDS) 20.51(MAP) 51.78(MeanIOU)	30.08(NDS) 20.46(MAP) 51.47(MeanIOU)	nuscenes det && seg	6x3x256x704, 10x128x128x2, 10x128x128x2	6.90	160.88
bev_gkt_mixvargenet_multitask_nuscenes	MixVarGENet	28.10(NDS) 19.91(MAP) 48.52(MeanIOU)	27.98(NDS) 19.99(MAP) 48.42(MeanIOU)	27.90(NDS) 20.00(MAP) 48.35(MeanIOU)	nuscenes det && seg	6x3x512x960, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2, 6x64x64x2	16.31	64.30
bev_ipm_4d_efficientnetb0_multitask_nuscenes	efficientnetb0	37.21(NDS) 22.00(MAP) 52.87(MeanIOU)	37.32(NDS) 22.21(MAP) 53.83(MeanIOU)	37.34(NDS) 22.15(MAP) 53.87(MeanIOU)	nuscenes det && seg	6x3x512x960, 6x128x128x2, 1x64x128x128, 1x128x128x2	9.83	108.87
detr3d_efficientnetb3_nuscenes	efficientnetb3	33.04(NDS) 27.52(MAP)	32.84(NDS) 27.14(MAP)	32.81(NDS) 27.06(MAP)	nuscenes det	6x3x512x1408	37.63	26.92
petr_efficientnetb3_nuscenes	efficientnetb3	37.65(NDS) 30.38(MAP)	37.26(NDS) 29.29(MAP)	37.40(NDS) 29.33(MAP)	nuscenes det	6x3x512x1408	77.10	13.05
bevformer_tiny_resnet50_detection_nuscenes	resnet50	37.12(NDS) 26.79(MAP)	37.16(NDS) 26.50(MAP)	37.15(NDS) 26.59(MAP)	nuscenes det	6x3x480x800, 1x2500x256, 1x50x50x2, 6x20x32x2, 1x100x50x2, 6x640x4x2, 1x2500x1	28.29(J6M)	35.90(J6M)
bev_cft_efficientnetb3_nuscenes	efficientnetb3	32.79(NDS) 24.79(MAP)	32.50(NDS) 24.47(MAP)	32.42(NDS) 24.46(MAP)	nuscenes det	6x3x512x1408	36.50	27.75
bev_sparse_henet_tinym_nuscenes	henet_tinym	54.19(NDS) 43.38(MAP)	52.23(NDS) 42.17(MAP)	-	nuscenes det	6x3x256x704, 6x4x4, 1x384x11, 1x384x256	12.92(J6M)	79.68(J6M)

Keypoint Detection

HeatmapKeypointModel

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
keypoint_efficientnetb0_carfusion	efficientnetb0	94.33	94.30	94.31	carfusion	1x3x128x128	0.45	4550.72

Trajectory Prediction

DenseTNT

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
densetnt_vectornet_argoverse1	vectornet	1.2974	1.2989	1.3038	argoverse 1	30x9x19x32, 30x11x9x64, 30x1x1x96, 30x2x1x2048, 30x1x1x2048	11.30	144.725382

QCNet

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
qcnet_oe_argoverse2	-	80.09	79.54	78.21	argoverse 2	输入见下方list	3.62(J6M)	247.74

注解

qcnet_oe_argoverse2 的指标是 HitRate 这项。

qcnet_oe_argoverse2 模型输入shape为:

1x30x10, 1x10x30x30, 1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x1, 1x1x30x80, 1x1x30x80, 1x1x30x80, 1x1x30x6, 1x1x30x6, 1x1x30x6, 1x1x30x6, 1x1x30x30, 1x1x30x30, 1x1x30x30, 1x30x2x128, 1x30x6x128, 1x80, 1x1x80x80, 1x1x80x80, 1x1x80x80, 1x1x80x50, 1x1x80x50, 1x1x80x50, 1x80x80, 1x80x50, 1x30x30, 1x30x1, 1x80x80

Occupancy Prediction

FlashOcc

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
flashocc_henet_lss_occ3d_nuscenes	henet_tinym_imagenet	0.3674	0.3657	0.3693	occ3d_nuscenes	6x3x512x960, 10x128x128x2, 10x128x128x2	8.39(J6M)	119.23(J6M)

Online Map Construction

MapTROE

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
maptroe_henet_tinym_bevformer_nuscenes	henet_tinym_imagenet	0.6632	0.6577	0.6567	nuscenes	6x3x480x800, 1x1x50x100, 6x20x100x2, 1x100x100x2, 6x2000x4x2, 1x5000x1	11.03(J6M)	93.77(J6M)
maptroe_sparse_henet_tinym_nuscenes	henet_tinym_imagenet	0.5982	0.5999	0.5992	nuscenes	6x3x256x704, 6x4x4	13.22(J6M)	77.14(J6M)

Lidar Fusion

LidarFusion

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
bevfusion_pointpillar_henet_multisensor_multitask_nuscenes	henet_tinym_imagenet	64.28(NDS) 58.09(MAP) 51.77(MIOU)	62.91(NDS) 57.48(MAP) 52.51(MIOU)	-	nuscenes det && occ3d	1x5x20x40000, 40000x4, 6x3x512x960, 1x256x128x2, 6x5120x2x2, 1x16384x1	24.87(J6M)	41.21(J6M)
bev_sparse_lidar_fusion_henet_tinym_nuscenes	henet_tinym_imagenet	66.64(NDS) 61.16(MAP)	66.31(NDS), 60.70(mAP)	65.96(NDS) 60.35(mAP)	nuscenes det	6x3x256x704, 6x4x4, 1x384x11, 1x384x256, 1x5x20x40000, 40000x4	19.16(J6M)	52.20(J6M)

Bev Multitask

BevSparseMultitask

network	backbone	float	qat	quantization	dataset	input shape	bpu latency (ms)	FPS
bev_sparse_det_maptr_flashocc_henet_tinym_nuscenes	henet_tinym_imagenet	52.34(det NDS) 41.33(det mAP) 59.58(map mAP) 31.89 (occ, MIOU)	52.14(det NDS) 41.00(det mAP) 59.06(map mAP) 33.19 (occ, MIOU)	-	nuscenes det && map && occ3d	6x3x256x704, 6x4x4, 1x384x11, 1x384x256, 10x128x128x2, 10x128x128x2	28.38(J6M)	35.18(J6M)

#模型仓库modelzoo

#Classification

#Detection

#Segmentation

#Lidar

#Lane Detection

#Multiple Object Track

#Binocular depth estimation

#Bev

#Keypoint Detection

#Trajectory Prediction

#Occupancy Prediction

#Online Map Construction

#Lidar Fusion

#Bev Multitask