Files

Chengfang Lu e72bc061c5 feat: HSAP platform v2 — modular navigation, quality review, audit log, world model simulation

Major changes:
- New frontend (platform/web/): Vite + React 18 + TypeScript + Tailwind
- 4-module navigation: 数据送标 / 模型管理 / 车队管理 / 系统管理
- Data catalog with charts (DMS/ADAS/Lane 3-tab view)
- Quality review workflow (标注质检): Good/Fine/Bad scoring with auto-advance
- Audit enhancements: batch operations, rejection categories, Feishu notifications
- Operation audit log (操作日志)
- World model simulation studio (仿真工坊)
- Dataset version management with snapshots and diff
- ADAS 7-class dataset integration (138K images organized + compressed)
- User management with Feishu integration and pagination
- CRUD/search/filter on all pages, card layout redesign
- PIL-optimized image overlay rendering
- Auto-snapshot on build, in_review workflow stage
- Removed embedded algorithm code (now in workspace)

2026-06-03 11:40:21 +08:00

6.1 KiB

Raw Blame History

Welcome to PytorchAutoDrive benchmark

The current benchmark's FLOPs & Param count is entirely based on thop to identify underlying basic ops, which might be inaccurate. But FLOPs count is an estimate to begin with. What we are doing here, is simply providing a relatively fair benchmark for comparing different methods.

Lane detection performance

method	backbone	resolution	FPS	FLOPS(G)	Params(M)
Baseline	VGG16	360 x 640	56.36	214.50	20.37
Baseline	ResNet18	360 x 640	148.59	85.24	12.04
Baseline	ResNet34	360 x 640	79.97	159.60	22.15
Baseline	ResNet50	360 x 640	50.58	177.62	24.57
Baseline	ResNet101	360 x 640	27.41	314.36	43.56
Baseline	ERFNet	360 x 640	85.87	26.32	2.67
Baseline	ENet	360 x 640	56.63	4.26	0.95
Baseline	MobileNetV2	360 x 640	126.54	4.49	2.06
Baseline	MobileNetV3-Large	360 x 640	104.34	3.63	3.30
SCNN	VGG16	360 x 640	21.18	218.64	20.96
SCNN	ResNet18	360 x 640	21.12	89.38	12.63
SCNN	ResNet34	360 x 640	20.77	163.74	22.74
SCNN	ResNet50	360 x 640	19.59	181.76	25.16
SCNN	ResNet101	360 x 640	13.50	318.50	44.15
SCNN	ERFNet	360 x 640	18.40	30.46	3.26
LSTR	ResNet18s	360 x 640	98.13	1.15	0.77
LSTR	ResNet18s-2x	360 x 640	97.27	4.05	3.05
LSTR	ResNet18s	1080 x 1920	91.23	10.20	0.77
LSTR	ResNet18s	2160 x 4320	23.60	40.75	0.77
LSTR	ResNet34	360 x 640	63.52	34.54	22.34
RESA	ResNet18	360 x 640	67.66	61.35	6.61
RESA	ResNet34	360 x 640	54.49	101.74	11.99
RESA	ResNet50	360 x 640	44.80	105.71	12.46
RESA	ResNet101	360 x 640	25.14	242.45	31.46
RESA	MobileNetV2	360 x 640	60.53	12.80	4.63
RESA	MobileNetV3-Large	360 x 640	54.39	11.95	5.88
LaneATT	ResNet18	360 x 640	198.29	18.67	12.02
LaneATT	ResNet34	360 x 640	133.84	36.01	22.12
BézierLaneNet	ResNet18	360 x 640	212.83	14.77	4.10
BézierLaneNet	ResNet34	360 x 640	149.52	29.85	9.49
Baseline	VGG16	288 x 800	55.31	214.50	20.15
Baseline	ResNet18	288 x 800	136.28	85.22	11.82
Baseline	ResNet34	288 x 800	72.42	159.60	21.93
Baseline	ResNet50	288 x 800	49.41	177.60	24.35
Baseline	ResNet101	288 x 800	27.19	314.34	43.34
Baseline	ERFNet	288 x 800	88.76	26.26	2.68
Baseline	ENet	288 x 800	57.99	4.12	0.96
Baseline	MobileNetV2	288 x 800	129.24	4.41	2.00
Baseline	MobileNetV3-Large	288 x 800	107.83	3.56	3.25
Baseline	RepVGG-A0	288 x 800	162.61	207.81	9.06
Baseline	RepVGG-A1	288 x 800	117.30	339.83	13.54
Baseline	RepVGG-B0	288 x 800	103.68	390.83	15.09
Baseline	RepVGG-B1g2	288 x 800	36.91	1166.76	42.20
Baseline	RepVGG-B2	288 x 800	18.98	2310.13	81.23
Baseline	Swin-Tiny	288 x 800	51.90	44.24	27.72
SCNN	VGG16	288 x 800	21.40	218.62	20.74
SCNN	ResNet18	288 x 800	20.80	89.34	12.42
SCNN	ResNet34	288 x 800	19.77	163.72	22.52
SCNN	ResNet50	288 x 800	18.88	181.72	24.94
SCNN	ResNet101	288 x 800	13.42	318.46	43.94
SCNN	ERFNet	288 x 800	18.80	30.40	3.27
SCNN	RepVGG-A1	288 x 800	20.53	343.96	14.13
RESA	ResNet18	288 x 800	69.58	61.33	6.62
RESA	ResNet34	288 x 800	55.61	101.72	12.01
RESA	ResNet50	288 x 800	46.75	105.70	12.48
RESA	ResNet101	288 x 800	26.08	242.44	31.47
RESA	MobileNetV2	288 x 800	59.49	12.55	4.63
RESA	MobileNetV3-Large	288 x 800	53.85	11.70	5.88
LSTR	ResNet34	288 x 800	65.39	33.86	22.34
BézierLaneNet	ResNet18	288 x 800	210.79	14.66	4.10
BézierLaneNet	ResNet34	288 x 800	144.65	29.54	9.49

Segmentation performance:

method	resolution	FPS	FLOPS(G)	Params(M)
FCN	256 x 512	43.32	216.42	51.95
FCN	512 x 1024	12.06	865.69	51.95
FCN	1024 x 2048	3.06	3462.77	51.95
ERFNet	256 x 512	91.20	15.03	2.07
ERFNet	512 x 1024	85.51	60.11	2.07
ERFNet	1024 x 2048	21.53	240.44	2.07
ENet	256 x 512	59.31	2.72	0.35
ENet	512 x 1024	55.69	10.88	0.35
ENet	1024 x 2048	30.88	43.53	0.35
DeeplabV2	256 x 512	44.87	180.59	43.90
DeeplabV2	512 x 1024	12.93	722.37	43.90
DeeplabV2	1024 x 2048	3.23	2889.49	43.90
DeeplabV3	256 x 512	35.26	241.65	58.63
DeeplabV3	512 x 1024	10.26	966.61	58.63
DeeplabV3	1024 x 2048	2.56	3866.45	58.63

All results are the maximum value of 3 times on a RTX 2080Ti.

Lane detection post-processing are not counted.

LaneATT NMS is not counted yet.

Profiling Models Yourself

In the setting of mode=simple, we employ a random tensor to replace the real image. Therefore, we can avoid using the DataLoader to obtain the best performance of models.

This is also the setting for the above benchmark.

python tools/profiling.py --mode=simple \
                          --config=<config file path> \           
                          --times=3 \
                          --height=<image height in pixels> \
                          --width=<image width in pixels>

Same config mechanism and commandline overwrite by --cfg-options as in training/testing.

In the setting of mode=real, so as to simulate that the real camera transmit frames to models, we set 'batch_size=1' and 'num_workers=0' in the DataLoader. Just use --mode=real and probably provide an actual model by --checkpoint.

For detailed instructions and commandline shortcuts available, run:

python tools/profiling.py --help

6.1 KiB Raw Blame History

Welcome to PytorchAutoDrive benchmark

Lane detection performance

Segmentation performance:

Profiling Models Yourself

6.1 KiB

Raw Blame History