fix(detector): 移除 ongoing_actions/visible_entities 空字段检测 + 新增 repair sh

抽检确认这两个字段为空是合法内容状态(静物/黑帧/模糊帧),
VLM 重修也修不好,保留会导致断点续跑死循环。
L3 空字段检测缩减为 frame_summary + spatial_layout。
This commit is contained in:
2026-07-09 00:54:27 -04:00
parent 86573735aa
commit 8c383354c5
3 changed files with 38 additions and 14 deletions
+27
View File
@@ -0,0 +1,27 @@
#!/usr/bin/env bash
# 树修复管线:检测 + VLM 重生成 + 校验 + Q&A 补全
#
# 用法:
# bash scripts/repair_trees.sh # 默认 16 并发续跑
# bash scripts/repair_trees.sh --dry-run # 仅检测不修复
# bash scripts/repair_trees.sh --reaggregate-all # 强制全量重聚合
# CONCURRENCY=8 bash scripts/repair_trees.sh # 自定义并发数
#
# 日志输出:
# stderr → 终端实时显示
# logs/repair_trees.log → 完整日志(自动 rotation 50MB
# logs/repair_telemetry.db → LLM/VLM 调用遥测
# logs/repair_progress.json → 断点续跑进度
set -euo pipefail
CONCURRENCY="${CONCURRENCY:-16}"
conda activate Video-Tree-TRM
python tools/repair_trees.py \
--videos-dir store/videos \
--srt-dir data/Video-MME/subtitle \
--questions-dir store/questions/benchmarks/Video-MME \
--concurrency "$CONCURRENCY" \
"$@"