OV-COCO 鲁棒性评估 (Robustness Evaluation)
评估目标检测模型在图像退化条件下的性能,专为 OV-COCO 数据集设计,支持 base_ap50、novel_ap50、all_ap50 指标。
目录结构
robustness_eval/
├── test_robustness_ovcoco.py # OV-COCO 鲁棒性测试脚本
├── merge_robustness_results.py # 结果合并与报告生成
├── run_clearclip_robustness.sh # ClearCLIP 8-GPU 并行测试
├── run_clipself_robustness.sh # CLIPSelf 8-GPU 并行测试
├── logs/ # 运行日志
└── results/ # 测试结果
├── clearclip/
└── clipself/
退化类型 (15 种 benchmark)
| 类别 | 退化类型 |
|---|---|
| Noise | gaussian_noise, shot_noise, impulse_noise |
| Blur | defocus_blur, glass_blur, motion_blur, zoom_blur |
| Weather | snow, frost, fog, brightness |
| Digital | contrast, elastic_transform, pixelate, jpeg_compression |
严重程度
1-5 级,数字越大退化越严重。共 75 个场景 (15 类型 × 5 级别)。
评估指标
OV-COCO 特有指标
- base_ap50: 已知类别 (48 类) 的 AP@IoU=0.50
- novel_ap50: 新类别 (17 类) 的 AP@IoU=0.50
- all_ap50: 所有类别 (65 类) 的 AP@IoU=0.50
鲁棒性指标
- P (Performance): 原始图像上的性能
- mPC (mean Performance under Corruption): 所有退化条件下的平均性能
- rPC (relative Performance under Corruption): mPC / P,衡量鲁棒性
使用方法
1. 运行鲁棒性测试
cd /mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/robustness_eval
# ClearCLIP (8 GPU 并行,后台运行)
bash run_clearclip_robustness.sh --nohup
# CLIPSelf (8 GPU 并行,后台运行)
bash run_clipself_robustness.sh --nohup
# 两个可以同时运行(GPU 显存足够时)
bash run_clearclip_robustness.sh --nohup && bash run_clipself_robustness.sh --nohup
2. 监控进度
# 查看日志
tail -f logs/gpu0_clearclip.log
# 检查完成数量 (预期 75 个)
ls results/clearclip/*_results.pkl 2>/dev/null | wc -l
ls results/clipself/*_results.pkl 2>/dev/null | wc -l
3. 合并结果并生成报告
# ClearCLIP
python3 merge_robustness_results.py \
--results-dir results/clearclip \
--model-name ClearCLIP
# CLIPSelf
python3 merge_robustness_results.py \
--results-dir results/clipself \
--model-name CLIPSelf
4. 输出文件
- 日志报告: 控制台输出 base_ap50、novel_ap50、all_ap50 等汇总
- Excel 报告:
results/<model>/robustness_report.xlsx- Summary: P、mPC、rPC 汇总
- Category mPC: 按退化类别 (noise/blur/weather/digital) 统计
- Corruption Avg: 每种退化类型的平均值
- Base Ap50 / Novel Ap50 / All Ap50: 详细 15×5 矩阵
- Full Matrix: 完整结果(便于复制到论文)
单场景测试
如需测试单个场景:
cd /mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/CLIPSelf/F-ViT
export PYTHONPATH=$PWD:/opt/tiger/xiaomoguhzz/mmdetection:$PYTHONPATH
python3 /mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/robustness_eval/test_robustness_ovcoco.py \
configs/declip/fvit_vitb16_upsample_fpn_bs64_3e_ovcoco_clearclip.py \
work_dirs/clearclip_ovcoco/epoch_3.pth \
--out /tmp/test.pkl \
--corruptions gaussian_noise \
--severities 1 \
--eval bbox
注意事项
依赖安装:
pip install imagecorruptions openpyxl pandas运行时间: 75 个场景 × 4836 图像,8 GPU 并行约需 2-3 小时
与旧脚本的区别:
- 旧脚本使用
mmdet/test_robustness.py,只输出标准 COCO 指标 - 新脚本
test_robustness_ovcoco.py调用CocoDatasetOV.evaluate(),输出 OV-COCO 特有的 base/novel AP50
- 旧脚本使用
模型路径
| 模型 | Config | Checkpoint |
|---|---|---|
| ClearCLIP | configs/declip/fvit_vitb16_upsample_fpn_bs64_3e_ovcoco_clearclip.py |
work_dirs/clearclip_ovcoco/epoch_3.pth |
| CLIPSelf | configs/ov_coco/fvit_vitb16_upsample_fpn_bs64_3e_ovcoco_eva_clipself_proposals.py |
/opt/tiger/xiaomoguhzz/fvit_eva_vitb16_ovcoco_clipself_proposals.pth |