vault backup: 2026-04-24 13:13:42

This commit is contained in:
2026-04-24 13:13:42 +08:00
parent 15d6b14eda
commit 537ebea129
2 changed files with 16 additions and 5 deletions

View File

@@ -4,8 +4,8 @@
"title": "Design ComfyUI video segmentation and tracking",
"titleGenerationStatus": "success",
"createdAt": 1776871296744,
"updatedAt": 1777005991920,
"lastResponseAt": 1777005991920,
"updatedAt": 1777007031552,
"lastResponseAt": 1777007031552,
"sessionId": "50e1c568-9493-4c86-8ca1-1f50b0f2e724",
"providerState": {
"providerSessionId": "50e1c568-9493-4c86-8ca1-1f50b0f2e724"
@@ -15,12 +15,12 @@
],
"usage": {
"model": "kimi-for-coding",
"inputTokens": 86018,
"inputTokens": 88635,
"cacheCreationInputTokens": 0,
"cacheReadInputTokens": 0,
"contextWindow": 200000,
"contextTokens": 86018,
"percentage": 43,
"contextTokens": 88635,
"percentage": 44,
"contextWindowIsAuthoritative": true
}
}

View File

@@ -172,6 +172,17 @@ tags:
3. 参数调整(可选):
- `points_per_side`: 网格密度(默认 32越大越精细但越慢
- `pred_iou_thresh`: IOU 阈值过滤低质量 mask
- `stability_score_thresh`: 稳定性阈值(默认 0.95
**医疗/低对比度场景调优建议**
在处理医疗手术图像等低对比度、边界模糊场景时Auto Segmentation 容易因过滤过严而产生"角落碎块"(只保留高对比度杂色,漏掉目标器官)。遇到此现象时请重点调整以下两个参数:
| 参数 | 当前问题 | 建议修改 |
|:---|:---|:---|
| `stability_score_thresh` | 默认 0.95 要求过高。手术灯光反光、组织颜色接近AI 很难在全图点阵下获得 95% 稳定性的遮罩,导致宁缺毋滥、滤掉模糊器官 | **降至 0.85 ~ 0.90** |
| `points_per_side` | 默认 16 意味着全图只撒了 256 个探测点。对于 2560×1440 等高分辨率医疗画面,点阵空隙太大,容易漏掉细小器械或解剖结构 | **提高至 32 或 64** |
4. 输出:`mask``segmented_image``bbox`
**此时你可以先点击「运行」测试单帧分割效果。**