AI EVALUATION, IN CONTEXT.BY EQUALITY MACHINE

Model reports

WorkIng

在高速公路施工区识别障碍并规划轨迹。先区分真实道路与仿真,再看仿真训练是否真的能迁移到真实测试集。

2026 年 10 月 30 日提交的原论文 v1;11 月 3 日核查新增。主成绩限定固定的真实白天测试集,其他场景与任务分表展示。

施工区改变了什么?

临时车道、障碍和路牌同时考验感知与规划。WorkIng-Highway 使用六摄像头、128 线激光雷达和 RTK-GPS;WorkIng-CARLA 提供程序化施工场景。

WorkIng original paper v1

区分场景与数据划分

真实白天与夜间序列按不重叠子序列进行 80/10/10 训练、验证、测试划分,CARLA 留出城镇。夜间排除稀疏类别,仿真成绩也不对应同一评测样本群。

WorkIng original paper v1 · WorkIng · Table 4

主成绩使用一个明确范围

主成绩采用表 5 全部 15 个检测器/训练配置的白天 NDS。原文采用 0–100 刻度,且该版本省略 AAE。夜间、CARLA、检测误差和规划结果分别保留。

WorkIng original paper v1 · WorkIng · Table 5

先读训练数据标记

W 表示 WorkIng-Highway,D 表示 DrivIng,C 表示 WorkIng-CARLA。W+D+C 对仿真数据按场景分层采样;只用仿真训练不等于在仿真中测试。

WorkIng · Table 5 · WorkIng original paper v1

论文发布与可运行资源

论文已可读取。11 月 3 日核查时,作者 GitHub 仓库仅有 README 与 LICENSE,尚不能据此确认完整数据集、运行器或复现配置已公开。

WorkIng · author repository (release placeholder) · WorkIng original paper v1

表 4 · 域内检测

训练和测试均使用所标场景。

表 4 · 域内检测
SceneModelATE ↓ASE ↓AOE ↓mAP ↑NDS ↑
DayCenterPoint0.210.180.070.7679.40
DayPETR0.870.270.240.3743.52
DayBEVFusion0.230.230.140.7576.59
NightCenterPoint0.140.150.070.7881.55
NightPETR0.870.180.060.3344.04
NightBEVFusion0.220.250.200.7877.71
CARLACenterPoint0.050.150.060.9794.65
CARLAPETR0.660.190.050.5661.35
CARLABEVFusion0.200.320.320.9385.12

Day、Night、CARLA 对应不同测试样本群。mAP 为小数,NDS 为乘 100 后的分数。

WorkIng · Table 4

表 5 · 白天测试与训练来源

五种训练配置均在该场景的同一真实测试集上评测。

表 5 · 白天测试与训练来源
ModelTrainingATE ↓ASE ↓AOE ↓mAP ↑NDS ↑
CenterPointW0.210.180.070.7679.40
CenterPointD0.220.240.310.4555.81
CenterPointC0.500.611.070.1318.37
CenterPointW+D0.140.180.180.7879.70
CenterPointW+D+C0.140.180.170.7980.75
PETRW0.870.270.240.3743.52
PETRD1.140.390.230.1122.37
PETRC1.140.521.050.014.32
PETRW+D0.660.190.200.3747.44
PETRW+D+C0.660.120.070.4152.27
BEVFusionW0.230.230.140.7576.59
BEVFusionD0.330.421.410.0815.70
BEVFusionC0.780.730.640.1419.05
BEVFusionW+D0.140.180.150.7981.33
BEVFusionW+D+C0.120.130.080.8384.98

仅白天 NDS 进入主成绩。W 行与表 4 重复,主成绩不重复计数。

WorkIng · Table 5

表 5 · 夜间测试与训练来源

五种训练配置均在该场景的同一真实测试集上评测。

表 5 · 夜间测试与训练来源
ModelTrainingATE ↓ASE ↓AOE ↓mAP ↑NDS ↑
CenterPointW0.140.150.070.7881.55
CenterPointD0.600.430.300.4548.90
CenterPointC0.690.670.450.2530.36
CenterPointW+D0.140.140.050.8385.53
CenterPointW+D+C0.130.120.040.8687.63
PETRW0.870.180.060.3344.04
PETRD1.090.450.590.0715.17
PETRC1.130.921.190.000.97
PETRW+D0.610.210.110.4050.70
PETRW+D+C0.570.260.150.4452.31
BEVFusionW0.220.250.200.7877.71
BEVFusionD0.650.660.910.2424.50
BEVFusionC0.760.700.590.1923.82
BEVFusionW+D0.160.190.140.8282.76
BEVFusionW+D+C0.150.140.110.8887.69

仅白天 NDS 进入主成绩。W 行与表 4 重复,主成绩不重复计数。

WorkIng · Table 5

表 6 · 独立 ROADWork 迁移测试

这是 ROADWork 上的二维掩码识别,不是 WorkIng 三维主指标。

表 6 · 独立 ROADWork 迁移测试
Training / provenanceModelAPAP50AP75APSAPMAPL
Zero-shot foundation modelsOpenSeeD2.95.82.51.32.84.0
Zero-shot foundation modelsDetic4.26.34.62.15.56.0
Zero-shot foundation modelsOMG-LLaVA0.20.30.20.00.10.8
Train: C (WorkIng-CARLA only)Mask R-CNN4.135.074.232.743.485.12
Train: C (WorkIng-CARLA only)Mask DINO5.658.825.052.494.467.44
Train: R (ROADWork only)Mask R-CNN30.2349.3032.4117.0932.7842.60
Train: R (ROADWork only)Mask DINO38.0155.7041.7619.2038.1251.27
Train: C+R (combined)Mask R-CNN35.5650.6534.4219.5833.7439.39
Train: C+R (combined)Mask DINO43.1559.7344.0121.0243.4747.54

均为掩码 AP(%)。三个零样本行引用自 ROADWork,WorkIng 未重新运行;其余行在每个架构内使用一致训练配方。

WorkIng · Table 6

表 7 · 真实日志上的开环规划

日志重放中的车辆不会响应新计划。两个 tiny 规划器均使用作者的摄像头与激光雷达改造版,并移除 HD 地图分支。

表 7 · 真实日志上的开环规划
SceneModelL2 0.5s (m)L2 1s (m)L2 2s (m)L2 avg (m)Collision 0.5s (%)Collision 1s (%)Collision 2s (%)Collision avg (%)
DayUniAD_tiny0.230.551.450.740.110.335.972.14
DayVAD_tiny0.180.431.250.620.140.345.131.87
NightUniAD_tiny0.370.701.630.900.850.884.842.19
NightVAD_tiny0.350.861.600.940.850.884.762.16

两类指标均越低越好;不能解释为反应式驾驶成功率。

WorkIng · Table 7 · WorkIng original paper v1

表 8 · CARLA 反应式规划

作者在 85 条路线上的单次闭环运行结果;论文未把 Driving Score 相差 0.02 解读为确定排名。

表 8 · CARLA 反应式规划
ModelL2 0.5s (m)L2 1s (m)L2 2s (m)L2 avg (m)Collision 0.5s (%)Collision 1s (%)Collision 2s (%)Collision avg (%)DS ↑RC ↑IP ↑
UniAD_tiny0.230.481.340.680.230.752.841.270.630.800.79
VAD_tiny0.250.541.190.660.080.541.960.860.650.760.85

DS 为驾驶得分,RC 为路线完成,IP 为违规惩罚因子。模型仅使用 WorkIng-CARLA 训练。

WorkIng · Table 8

表 9 · 规划向真实数据迁移

训练采用 W、C 或 W+C,真实测试集保持不变。

表 9 · 规划向真实数据迁移
TrainingUniAD day L2 (m)UniAD night L2 (m)UniAD day collision (%)UniAD night collision (%)VAD day L2 (m)VAD night L2 (m)VAD day collision (%)VAD night collision (%)
W0.740.902.142.190.620.941.872.16
C7.578.136.985.518.329.336.785.82
W+C0.660.841.171.650.530.831.091.78

数值平均了 0.5/1.0/2.0 秒三个时域,越低越好。

WorkIng · Table 9

表 12 · 分类检测结果

保留作者表中的每个类别,不把排除的类别改记为零。

表 12 · 分类检测结果
SceneModelCarTruckBusTrailerBarrierMotorcyclePedestrianTotal
DayCenterPoint0.950.890.370.870.860.680.710.76
DayPETR0.600.470.180.390.390.450.120.37
DayBEVFusion0.950.850.370.830.800.800.640.75
NightCenterPoint0.880.87-0.440.92--0.78
NightPETR0.430.57-0.010.30--0.33
NightBEVFusion0.930.85-0.420.90--0.78
CARLACenterPoint0.990.980.99-0.880.980.980.97
CARLAPETR0.640.520.53-0.650.550.490.56
CARLABEVFusion0.970.950.94-0.820.960.920.93

所有成绩为小数 mAP;横杠表示排除或缺失类别。

WorkIng · Table 12

原始发布

2026 年 10 月 30 日提交的原论文 v1;11 月 3 日核查新增。主成绩限定固定的真实白天测试集,其他场景与任务分表展示。

核对于 2026-11-03

已收录的评测资料

这条来源记录收录了 15 条数值成绩。来源区分了 3 个模型身份或配置。同一模型可能有多条记录;比较前请核对具体版本、工具和运行预算。 查看成绩及逐条证据 →

报告指标
WorkIng-Highway DAY test NDS
原始单位
points
指标方向
数值越高越好

从这些资料开始

引用文档及日期(1)

评测信息与证据

来源
Model reports
发布者
未收录
评测发布日期
2026-10-30
采集日期
未收录
模态
未收录

本站保留来源报告的结果,不代表独立复测。请先核对任务版本、评测协议、分数单位和报告日期;不同运行条件不能直接合并为统一排名。

报告成绩 15

按来源中的观察顺序展示,顺序不代表排名。日期保留原始精度。

模型数值单位协议报告日期与口径来源
CenterPoint18.37pointsTable 5; fixed WorkIng-Highway DAY test split; training=C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); LiDAR-only, single sweep, 100m x 100m grid, voxel (0.1,0.1,0.2)m, 25 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
BEVFusion19.05pointsTable 5; fixed WorkIng-Highway DAY test split; training=C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-LiDAR fusion, Swin-Tiny, VoxelNet voxel (0.075,0.075,0.2)m, 50 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
PETR4.32pointsTable 5; fixed WorkIng-Highway DAY test split; training=C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-only, V2-99-eSE, 384x960 input, 75 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
CenterPoint55.81pointsTable 5; fixed WorkIng-Highway DAY test split; training=D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); LiDAR-only, single sweep, 100m x 100m grid, voxel (0.1,0.1,0.2)m, 25 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
BEVFusion15.7pointsTable 5; fixed WorkIng-Highway DAY test split; training=D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-LiDAR fusion, Swin-Tiny, VoxelNet voxel (0.075,0.075,0.2)m, 50 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
PETR22.37pointsTable 5; fixed WorkIng-Highway DAY test split; training=D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-only, V2-99-eSE, 384x960 input, 75 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
CenterPoint79.4pointsTable 5; fixed WorkIng-Highway DAY test split; training=W (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); LiDAR-only, single sweep, 100m x 100m grid, voxel (0.1,0.1,0.2)m, 25 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
BEVFusion76.59pointsTable 5; fixed WorkIng-Highway DAY test split; training=W (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-LiDAR fusion, Swin-Tiny, VoxelNet voxel (0.075,0.075,0.2)m, 50 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
PETR43.52pointsTable 5; fixed WorkIng-Highway DAY test split; training=W (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-only, V2-99-eSE, 384x960 input, 75 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
CenterPoint79.7pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); LiDAR-only, single sweep, 100m x 100m grid, voxel (0.1,0.1,0.2)m, 25 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
BEVFusion81.33pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-LiDAR fusion, Swin-Tiny, VoxelNet voxel (0.075,0.075,0.2)m, 50 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
PETR47.44pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-only, V2-99-eSE, 384x960 input, 75 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
CenterPoint80.75pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D+C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); LiDAR-only, single sweep, 100m x 100m grid, voxel (0.1,0.1,0.2)m, 25 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
BEVFusion84.98pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D+C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-LiDAR fusion, Swin-Tiny, VoxelNet voxel (0.075,0.075,0.2)m, 50 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results
PETR52.27pointsTable 5; fixed WorkIng-Highway DAY test split; training=W+D+C (W=WorkIng-Highway; D=DrivIng; C=WorkIng-CARLA); camera-only, V2-99-eSE, 384x960 input, 75 epochs; nuScenes adapted NDS x100 excludes AAE; W+D+C samples 1/8 CARLA frames per scenario; not night or CARLA-test performance2026-10-30 · document_publicationWorkIng paper v1 · reported benchmark results