6.6 KiB

Raw Blame History

Phase 3 Epoch 23 Baseline性能报告

Checkpoint: runs/enhanced_from_epoch19/epoch_23.pth
配置: EnhancedBEVSegmentationHead, 400×400分辨率, 2层Decoder
数据来源: 训练日志validation结果
生成时间: 2025-10-30

📊 3D检测性能 (Epoch 23)

总体指标

NDS (nuScenes Detection Score): 0.6941  ⭐ 优秀
mAP (mean Average Precision):   0.6446  ⭐ 优秀

错误指标:
  mATE (Translation Error):     0.2829 m
  mASE (Scale Error):           0.2561
  mAOE (Orientation Error):     0.3098 rad
  mAVE (Velocity Error):        0.2468 m/s
  mAAE (Attribute Error):       0.1869

各类别AP (Average Precision)

Car (最重要类别)

AP @0.5m: 0.7662
AP @1.0m: 0.8664
AP @2.0m: 0.8926
AP @4.0m: 0.9039  ⭐ 优秀

Errors:
  Translation: 0.1853 m
  Scale: 0.1534
  Orientation: 0.0594 rad
  Velocity: 0.2523 m/s
  Attribute: 0.1881

Pedestrian

AP @0.5m: 0.8240
AP @1.0m: 0.8366
AP @2.0m: 0.8465
AP @4.0m: 0.8579  ⭐ 优秀

Errors:
  Translation: 0.1326 m
  Scale: 0.2875
  Orientation: 0.3548 rad
  Velocity: 0.2076 m/s
  Attribute: 0.0918

Truck

AP @0.5m: 0.3965
AP @1.0m: 0.5694
AP @2.0m: 0.6640
AP @4.0m: 0.7101

Errors:
  Translation: 0.3305 m
  Scale: 0.1835
  Orientation: 0.0617 rad
  Velocity: 0.2344 m/s
  Attribute: 0.2097

Bus

AP @0.5m: 0.4871
AP @1.0m: 0.7293
AP @2.0m: 0.8429
AP @4.0m: 0.8612  ⭐ 优秀

Errors:
  Translation: 0.3330 m
  Scale: 0.1841
  Orientation: 0.0446 rad
  Velocity: 0.4485 m/s
  Attribute: 0.2647

Construction Vehicle (困难类别)

AP @0.5m: 0.0386  ⚠️ 低
AP @1.0m: 0.2014
AP @2.0m: 0.3661
AP @4.0m: 0.4439

Errors:
  Translation: 0.7056 m  ⚠️ 高
  Scale: 0.4520  ⚠️ 高
  Orientation: 0.9092 rad  ⚠️ 高
  Velocity: 0.1110 m/s
  Attribute: 0.3167

Trailer

AP @0.5m: 0.1565
AP @1.0m: 0.3713
AP @2.0m: 0.5459
AP @4.0m: 0.6612

Errors:
  Translation: 0.4756 m
  Scale: 0.2077
  Orientation: 0.6311 rad
  Velocity: 0.2159 m/s
  Attribute: 0.1602

Barrier

AP @0.5m: 0.5786
AP @1.0m: 0.6772
AP @2.0m: 0.7162
AP @4.0m: 0.7304

Errors:
  Translation: 0.1882 m
  Scale: 0.2790
  Orientation: 0.0505 rad

Motorcycle

AP @0.5m: 0.6062
AP @1.0m: 0.7315
AP @2.0m: 0.7599
AP @4.0m: 0.7687

Errors:
  Translation: 0.1917 m
  Scale: 0.2397
  Orientation: 0.2711 rad
  Velocity: 0.3278 m/s
  Attribute: 0.2569

Bicycle

AP @0.5m: 0.5482
AP @1.0m: 0.5816
AP @2.0m: 0.5883
AP @4.0m: 0.6018

Errors:
  Translation: 0.1643 m
  Scale: 0.2557
  Orientation: 0.4058 rad
  Velocity: 0.1771 m/s
  Attribute: 0.0071

Traffic Cone

AP @0.5m: 0.7435
AP @1.0m: 0.7510
AP @2.0m: 0.7676
AP @4.0m: 0.7935

Errors:
  Translation: 0.1218 m
  Scale: 0.3181

📊 BEV分割性能 (Epoch 23)

总体指标

mean IoU @max: 0.4130  (41.3%)

各类别IoU @max (最佳阈值)

Drivable Area (最大类别)

IoU @max: 0.7063  ⭐ 优秀

不同阈值下的IoU:
  @0.35: 0.7045
  @0.40: 0.7063
  @0.45: 0.6964
  @0.50: 0.6770
  @0.55: 0.6504
  @0.60: 0.6201
  @0.65: 0.5880

Pedestrian Crossing

IoU @max: 0.3931  

不同阈值:
  @0.35: 0.3931
  @0.40: 0.3700
  @0.45: 0.3266
  @0.50: 0.2747
  @0.55: 0.2212
  @0.60: 0.1725
  @0.65: 0.1292

Walkway

IoU @max: 0.5278  ✅ 良好

不同阈值:
  @0.35: 0.5278
  @0.40: 0.5189
  @0.45: 0.4948
  @0.50: 0.4619
  @0.55: 0.4243
  @0.60: 0.3853
  @0.65: 0.3450

Stop Line (关键改进目标)

IoU @max: 0.2657  ⚠️ 需大幅提升

不同阈值:
  @0.35: 0.2657
  @0.40: 0.2245
  @0.45: 0.1787
  @0.50: 0.1372
  @0.55: 0.1018
  @0.60: 0.0732
  @0.65: 0.0506

问题: 细线特征，0.3m分辨率不足

Carpark Area

IoU @max: 0.3948

不同阈值:
  @0.35: 0.3948
  @0.40: 0.3758
  @0.45: 0.3492
  @0.50: 0.3193
  @0.55: 0.2877
  @0.60: 0.2559
  @0.65: 0.2230

Divider (关键改进目标)

IoU @max: 0.1903  ⚠️ 需大幅提升

不同阈值:
  @0.35: 0.1903
  @0.40: 0.1361
  @0.45: 0.0856
  @0.50: 0.0470
  @0.55: 0.0222
  @0.60: 0.0090
  @0.65: 0.0029

问题: 最细线特征，0.3m分辨率严重不足

🎯 Phase 3性能总结

优势 ✅

3D检测:
  - NDS 0.6941 (行业领先)
  - Car, Pedestrian, Motorcycle, Traffic Cone AP高
  
BEV分割:
  - Drivable Area 0.7063 (优秀)
  - Walkway 0.5278 (良好)

弱点 ⚠️

3D检测:
  - Construction Vehicle AP低 (0.04-0.44)
  - Trailer AP中等 (0.16-0.66)

BEV分割:
  - Stop Line IoU仅0.2657 (目标0.35+)
  - Divider IoU仅0.1903 (目标0.28+)
  - 整体mIoU 0.4130 (目标0.48+)

根本原因: 0.3m分辨率对细线类别不够

🎯 Stage 1改进目标

基于Epoch 23 baseline，Stage 1 (600×600)的目标:

BEV分割改进目标

Stop Line IoU:
  Epoch 23: 0.2657
  目标: 0.35+ (+31%)
  策略: 分辨率提升 + 4层Decoder + Deep Supervision

Divider IoU:
  Epoch 23: 0.1903
  目标: 0.28+ (+47%)
  策略: 同上

整体mIoU:
  Epoch 23: 0.4130
  目标: 0.48+ (+16%)

3D检测目标

NDS: 保持0.69+ (不下降)
mAP: 保持0.64+ (不下降)

理想: 小幅提升

📋 Epoch 1评估checklist

对比指标

3D检测 (vs Epoch 23):

NDS: ? vs 0.6941
mAP: ? vs 0.6446
各类别AP变化

BEV分割 (vs Epoch 23):

mIoU: ? vs 0.4130
Drivable Area: ? vs 0.7063
Ped Crossing: ? vs 0.3931
Walkway: ? vs 0.5278
Stop Line: ? vs 0.2657 ⭐ 关注
Carpark: ? vs 0.3948
Divider: ? vs 0.1903 ⭐ 关注

改进归因

分辨率提升的贡献
Deep Supervision的贡献
Dice Loss的贡献
4层Decoder的贡献

📊 预期性能对比表

指标	Epoch 23 (Phase 3)	Epoch 1目标	Epoch 10目标
3D检测
NDS	0.6941	0.69+	0.69+
mAP	0.6446	0.64+	0.64+
BEV分割
mIoU	0.4130	0.44+	0.48+
Drivable	0.7063	0.71+	0.72+
Ped Cross	0.3931	0.42+	0.45+
Walkway	0.5278	0.54+	0.56+
Stop Line ⭐	0.2657	0.30+	0.35+
Carpark	0.3948	0.42+	0.45+
Divider ⭐	0.1903	0.22+	0.28+

⏭️ 下一步

立即 (已完成)

✅ 提取Epoch 23 baseline数据
✅ 生成详细性能报告

Epoch 1后 (~21小时)

⏸️ 评估epoch_1.pth
⏸️ 对比Epoch 23 baseline
⏸️ 量化初步改进

Epoch 5后 (~4.5天)

⏸️ 评估epoch_5.pth
⏸️ 评估是否达到中期目标

Stage 1完成 (~9天)

⏸️ 最终性能评估
⏸️ 与Epoch 23全面对比
⏸️ 生成改进归因分析

Baseline已建立！可用于后续所有对比分析。

6.6 KiB Raw Blame History Unescape Escape

Phase 3 Epoch 23 Baseline性能报告

📊 3D检测性能 (Epoch 23)

总体指标

各类别AP (Average Precision)

Car (最重要类别)

Pedestrian

Truck

Bus

Construction Vehicle (困难类别)

Trailer

Barrier

Motorcycle

Bicycle

Traffic Cone

📊 BEV分割性能 (Epoch 23)

总体指标

各类别IoU @max (最佳阈值)

Drivable Area (最大类别)

Pedestrian Crossing

Walkway

Stop Line (关键改进目标)

Carpark Area

Divider (关键改进目标)

🎯 Phase 3性能总结

优势 ✅

弱点 ⚠️

🎯 Stage 1改进目标

BEV分割改进目标

3D检测目标

📋 Epoch 1评估checklist

对比指标

改进归因

📊 预期性能对比表

⏭️ 下一步

立即 (已完成)

Epoch 1后 (~21小时)

Epoch 5后 (~4.5天)

Stage 1完成 (~9天)

6.6 KiB

Raw Blame History