agent-l / docs /analysis /error_analysis_20260524.md
zhou777's picture
Add files using upload-large-folder tool
0525670 verified
|
Raw
History Blame Contribute Delete
9.54 kB

Error Analysis 2026-05-24

本文档汇总 A official、B full、B-lite 和 B old + occlusion_only 的同口径逐图错误分析。 当前目标是先判断语义信息在什么场景帮助主任务、在什么场景伤害主任务,再决定是否继续训练新模型。

输入运行

  • A_official: /home/yyao/lsy/agent-base-linux-codex/runs/validation_eval/qwen3_5_35b_a3b_exp_a_official
  • B_full_rerun: /home/yyao/lsy/agent-base-linux-codex/runs/validation_eval/qwen3_5_35b_a3b_exp_b_attr_rerun_20260501_174649_max12000_stream
  • B_old_occlusion_only: /home/yyao/lsy/agent-base-linux-codex/runs/validation_eval/qwen3_5_35b_a3b_ablation_occlusion_only_tp4_20260501_2058
  • B_lite_dataset: /home/yyao/lsy/agent-base-linux-codex/runs/validation_eval/qwen3_5_35b_a3b_exp_b_lite_v0_20260504_2206_dataset

总体指标

Run Precision Recall F1 Avg IoU Maturity Acc Occlusion Acc Pred Matched Risk Cases Mean Risk
A_official 0.9145 0.8857 0.8999 0.9074 0.9408 0.0 1076 984 20 1.5859
B_full_rerun 0.8668 0.9253 0.8951 0.9033 0.8998 0.7463 1186 1028 20 1.8125
B_old_occlusion_only 0.9046 0.9046 0.9046 0.906 0.9107 0.6975 1111 1005 20 1.5547
B_lite_dataset 0.876 0.9091 0.8922 0.904 0.8909 0.7351 1153 1010 20 1.875

初步结论

  • 当前 F1 最好的口径是 B_old_occlusion_only,F1 为 0.9046
  • 当前遮挡属性最好的口径是 B_full_rerun,occlusion acc 为 0.7463
  • 当前成熟度准确率最高的是 A_official,maturity acc 为 0.9408
  • B-lite 能学习遮挡属性,但没有超过 A official 的检测和成熟度基线,说明语义信息不宜继续以长文本输出形式直接压到检测任务上。
  • 下一阶段应优先把语义用于后验验证、风险筛选和伪标签过滤。

错误标签分布

Error Tag Image Count
bbox_alignment_risk 14
dense_scene_sensitive 36
maturity_error_risk 61
occlusion_error_risk 77
possible_false_positive 71
possible_missed_detection 67

Risk Case 重叠

Run Risk Case Count
A_official 20
B_full_rerun 20
B_old_occlusion_only 20
B_lite_dataset 20

Risk case 被多个实验同时命中的分布:

命中实验数 图像数
1 10
2 12
3 10
4 4

最高风险样本

Image Key Max Risk Risk Runs Error Tags Best F1 Worst F1
IMG_20191215_112843_jpg.rf.0e020631b2ff9bf178eefcc9e309b875_orig.jpg 8 4 bbox_alignment_risk, dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun B_lite_dataset
IMG_1187_jpg.rf.3d3e2bc9c31942b052ed4cb4889546cd_orig.jpg 7 4 bbox_alignment_risk, dense_scene_sensitive, possible_false_positive, possible_missed_detection B_old_occlusion_only A_official
IMG_1242_jpg.rf.26200c8f2c2a985094bb5d3c03b5e754_orig.jpg 7 4 dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_lite_dataset
IMG_20191215_112604_jpg.rf.818f95a56efa2cb722edacc3af26c09d_orig.jpg 7 4 bbox_alignment_risk, dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_full_rerun
IMG_20191215_112836_jpg.rf.8a499c79911328bb6879fa58c10768ed_orig.jpg 7 4 dense_scene_sensitive, possible_false_positive, possible_missed_detection B_full_rerun A_official
IMG_20191215_112909.jpg 7 4 dense_scene_sensitive, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_old_occlusion_only
IMG20211230170257_00_jpg.rf.9c0fcbc6329d4bc4bb29a19b21ffc3d5.jpg 7 3 maturity_error_risk, possible_false_positive, possible_missed_detection B_lite_dataset B_full_rerun
IMG20211226161001_00_jpg.rf.c7c10a7d94dd5b7df7bd0950509d592d.jpg 6 4 maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_old_occlusion_only
IMG20211226162202_00_jpg.rf.9aa4e47e9ee261e02716a19550c8a344.jpg 6 4 occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_full_rerun
IMG_20191215_112617_jpg.rf.b1a388efd223e03fec31d3df9f5f437b.jpg 6 4 possible_false_positive, possible_missed_detection B_full_rerun A_official
IMG20211226160908_00_jpg.rf.8c697b77eb53b1c176486e3a1fb26f14.jpg 6 3 bbox_alignment_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun B_lite_dataset
IMG20211226163021_00_jpg.rf.7b50dd7bb9ffd12705c288a75a722823.jpg 5 4 maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_full_rerun
IMG20211230174505_00_jpg.rf.96444d7eb331967fac99820ddce4fbaa.jpg 5 4 dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun B_old_occlusion_only
IMG_1130_jpg.rf.e742c27f51bcd188352761cc173d0b8a_orig.jpg 5 4 dense_scene_sensitive, maturity_error_risk, possible_false_positive, possible_missed_detection B_lite_dataset A_official
IMG_1139_jpg.rf.8757303148dc7b66106fd3e641459fd8_orig.jpg 5 4 bbox_alignment_risk, dense_scene_sensitive, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun B_old_occlusion_only
IMG_20191215_111017_jpg.rf.83704e10824aba4c89b1854ea5f73fd3.jpg 5 4 dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun A_official
IMG_20191215_111513_jpg.rf.f2e996d41716becea631c4d684f7dd7b.jpg 5 4 maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_full_rerun
IMG_20191215_111949_jpg.rf.87f750340c92125f381ac94487d8a9ff.jpg 5 4 bbox_alignment_risk, dense_scene_sensitive, occlusion_error_risk, possible_false_positive, possible_missed_detection B_full_rerun A_official
IMG_20191215_112325.jpg 5 4 dense_scene_sensitive, maturity_error_risk, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_old_occlusion_only
IMG_20191215_112339_jpg.rf.f2737ec8f97d2af58f42249a0b6a4f73_orig.jpg 5 4 dense_scene_sensitive, occlusion_error_risk, possible_false_positive, possible_missed_detection A_official B_lite_dataset

人工复核图

已导出最高风险 Top 30 样本的五联对比图:

  • 输出目录:runs/analysis/validation_error_analysis_20260524/visual_review_images/
  • 高清 300 DPI 输出目录:runs/analysis/validation_error_analysis_20260524/visual_review_images_300dpi/
  • 说明文件:runs/analysis/validation_error_analysis_20260524/visual_review_images/README.md
  • manifest:runs/analysis/validation_error_analysis_20260524/visual_review_images/manifest.jsonl

每张图包含:

  1. GT。
  2. A official。
  3. B full rerun。
  4. B old + occlusion_only
  5. B-lite dataset。

建议优先查看前 10 张,它们在多个实验中都被标为高风险,适合判断密集目标、误检、漏检和成熟度错误的主要来源。

人工复核结果

人工复核记录见:

  • docs/reviews/people/validation_error_analysis_20260524.md
  • docs/reviews/people/validation_error_analysis_20260524_summary.md

人工复核前 10 张最高风险图后,当前结论需要修正:

  1. 多数高风险样本中的 GT 存在明显漏标,尤其是密集、遮挡和小目标番茄。
  2. 模型多出来的框基本没有明显多框或重复框问题,很多更像是真实番茄候选。
  3. B-lite 在前 5 张高风险密集图中最接近人工判断。
  4. A official 在部分图像上仍更稳,尤其 006、007、008、010。
  5. 当前自动评估中的 false positive 可能包含大量 GT 漏标候选,因此 precision 可能被低估。

因此,下一步应优先推进漏标候选挖掘和 Gold v1 重审,而不是直接把模型多框解释为误检。

漏标候选挖掘结果

已基于 A official、B full rerun、B old + occlusion_only、B-lite dataset 的预测结果生成漏标候选:

  • 脚本:pipelines/mine_missing_label_candidates.py
  • 文档:docs/analysis/missing_label_candidates_20260524.md
  • 输出目录:runs/analysis/missing_label_candidates_20260524/

结果摘要:

  • 原始 unmatched prediction:449
  • 聚类后候选:227
  • 含候选图像数:64
  • 四模型共同支持候选:22
  • 三模型共同支持候选:46

下一步

  1. 人工优先查看 runs/analysis/missing_label_candidates_20260524/preview_images/ 中 support=4 和 support=3 的候选。
  2. 建立 Gold v1 Review 层,保留 Gold v1 不变,避免破坏数据可追溯性。
  3. SemanticVerifier 对候选框做番茄性、成熟度和遮挡验证。
  4. 基于 accepted 候选构建 Gold v2 Candidate。
  5. 在 review-adjusted GT 形成前,不建议继续训练新的长文本属性增强模型。