/root/.openclaw/workspace/tasks/task-keyence-r4a/task.md/root/.openclaw/workspace/tasks/task-keyence-r4a/session.md对「旋转目标检测(Rotated / Oriented Object Detection)」领域做广度搜索,快速摸清主要算法、数据集、开源实现,以及是否存在适合少样本+CPU场景的旋转检测方案,为 R4b 深度研究提供输入。
关注重点:
- 旋转边界框检测算法(OBB - Oriented Bounding Box)
- 有没有支持少样本(1-5张)的旋转目标检测
- CPU 推理可行性
researcher → 爱衣质检
agent:researcher:main)任务:对旋转目标检测领域进行广度搜索(Search模式),快速建立领域地图。
搜索目标(必须覆盖以下维度):
1. 主流算法:
- RRPN(Rotated Region Proposal Network)
- RoI-Transformer / Oriented RCNN
- CSL(Circular Smooth Label)
- SCRDet / SCRDet++
- YOLOv8-OBB / YOLO-OBB 系列
- S2A-Net、LSKNet、FCOSR
- RetinaNet-OBB、Gliding Vertex
2. 数据集:DOTA、DOTA-v2、HRSC2016、UCAS-AOD、FAIR1M、SKU110K-R
3. 开源框架:mmrotate、OBBDetection、ultralytics OBB
4. 少样本+旋转:有没有论文研究 few-shot rotated object detection?
5. 轻量化旋转检测:CPU上可运行的旋转检测(ONNX/TFLite/量化)?
搜索策略(Search模式):
- SearXNG 广度搜索:8条查询
- "rotated object detection" survey 2024 2025
- oriented bounding box detection deep learning OBB
- "few-shot rotated object detection" OR "few-shot oriented detection"
- YOLOv8 OBB rotated detection CPU inference
- mmrotate oriented RCNN DOTA benchmark
- rotated object detection lightweight mobile CPU
- 旋转目标检测 综述 算法对比
- rotated bounding box detection few training samples
- search.py 定向:
- arxiv × 10(oriented object detection survey)
- openalex × 8(rotated bounding box detection)
- github × 10(rotated object detection OBB)
- hackernews × 5(rotated detection)
- top 5-8 页面提取正文
输出要求:
- 报告路径:/root/.openclaw/workspace-researcher/reports/rotated-object-detection-breadth-survey.md
- 执行日志:追加到 /root/.openclaw/workspace/tasks/task-keyence-r4a/session.md
- freshness_type: academic
- 来源数量目标:≥ 25 条
开始时:
1. 发工作日志:
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh researcher receive "旋转目标检测广度搜索" task-keyence-r4a
完成后:
1. 将执行日志追加到 session.md
2. 发工作日志:
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh researcher handoff "旋转目标检测广度搜索" main task-keyence-r4a
3. sessions_send 通知爱衣(agent:main:main,必须传 timeoutSeconds=0,禁止省略):
task_id=task-keyence-r4a
task=/root/.openclaw/workspace/tasks/task-keyence-r4a/task.md
du -sb /root/.openclaw/workspace/tasks/task-keyence-r4a/
wc -l /root/.openclaw/workspace/tasks/task-keyence-r4a/session.md
若行数 N > 0,read session.md 全文。
任务特定检查:
- 是否覆盖了至少 5 种旋转检测算法(含YOLO-OBB系、Oriented RCNN系)?
- 是否有关于「少样本+旋转」的搜索结果(哪怕是"该细分领域论文极少"也要写明)?
- 是否有开源框架的信息(mmrotate 等)?
- 是否有 CPU 可行性的初步判断?
- 来源数量是否 ≥ 20 条?
通过 →
1. 立即启动 task-keyence-r4b:
- 重置 researcher session:
bash
python3 -c "
import json
f='/root/.openclaw/agents/researcher/sessions/sessions.json'
d=json.load(open(f)); d.pop('agent:researcher:main',None); json.dump(d,open(f,'w'))
"
- 占锁 task-keyence-r4b:
bash
mkdir /root/.openclaw/workspace/tasks/task-keyence-r4b/.lock 2>/dev/null
- 发链路日志:
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh main start "旋转目标检测深度研究" "researcher → main" task-keyence-r4b
- sessions_send 给 researcher(timeoutSeconds=0):
task_id=task-keyence-r4b
task=/root/.openclaw/workspace/tasks/task-keyence-r4b/task.md
2. 发工作日志(R4a完成):
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh main done "旋转目标检测广度搜索" task-keyence-r4a
不通过(rejectCount == 0) →
1. 分析问题,创建 task-keyence-r4a-retry1
2. 在原 session.md 末尾追加 rejectCount=1
3. sessions_send 给 researcher
4. 发工作日志:
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh main retry "旋转目标检测广度搜索" "researcher → main" researcher 1 task-keyence-r4a
rejectCount >= 1 →
1. 发工作日志:
bash
/root/.openclaw/workspace/scripts/log-to-channel.sh main fail "旋转目标检测广度搜索" task-keyence-r4a
2. message 主人,请裁决