Unify FreiHAND shared-v1.0 batch 3

FreiHAND/shared-v1.0/doc/validation/REPORT.md

@@ -0,0 +1,58 @@
+# FreiHAND 核验报告（小样本人工确认通过）
+
+[打开离线核验页面](index.html) · [核验范围与文件摘要](evidence.json)
+
+## 当前结论
+
+- 小样本转换自动检查通过；15 张预览的 mask 已与 raw 解码数值逐像素核对。
+- **小样本人工确认：用户于 2026-09-18T16:00:21 确认通过。全量自动验收及额外随机抽查已通过，详见下文；不代表对每张图都做过人工检查。**
+- 用户已确认当前 15 个样本没有问题，允许继续全量转换；全量完成后补充本地验收结果，再决定发布。
+
+## 范围和检查方法
+
+12 张训练图覆盖三个基础样本的全部四种背景版本，另含三张评估图。选择是有目的的抽查，不是随机统计质量估计。FreiHAND 为独立图片，这里不验证视频时序。
+
+自动检查包括 RGB 原 JPEG 码流及固定解码路径的像素一致性、分割值一致性、规范数组序列化和 MANO 根位置恢复。页面依次展示 RGB、分割叠加、源 XYZ 投影与转换 MANO 投影；可单独打开原 JPEG。人工检查关注图片/标注对应、左右手、方向、尺度、投影与分割边缘。
+
+## 已知问题与边界
+
+预览曾错误地让 FFmpeg 将 gray8 扩展为 gray16，强度被放大 257 倍，造成荧光绿色叠加。已改为按存储位深解码，再在数组中拓宽类型；15 个预览重新核对通过。该问题在预览读取端，保存的 mask 未因此改变。
+
+MANO 拟合与原生 XYZ 允许存在源标注残差，不能要求完全重合。训练 JPEG mask 保留灰度，不二值化；评估区域图仅确认手/背景类别，不猜测细分手指名称。源 JSON 到规范 float32 的舍入与视频无损编码是两件事。
+
+本报告只绑定 evidence.json 中的样本产物和脚本摘要；后续全量产物需另行验收，不能直接继承“通过”。HTML 和配图均为相对链接，下载整个 validation 目录即可离线查看，无需原始包、Python 或模型权重。配图仅为核验展示，不替代 observation 中的视频数据。
+
+## 样本清单
+
+| 样本 ID | 原标注索引 | 背景版本 |
+|---|---:|---:|
+| FreiHAND/evaluation/00000000 | 0 | 0 |
+| FreiHAND/evaluation/00001000 | 1000 | 0 |
+| FreiHAND/evaluation/00003959 | 3959 | 0 |
+| FreiHAND/training/00000000 | 0 | 0 |
+| FreiHAND/training/00001000 | 1000 | 0 |
+| FreiHAND/training/00030000 | 30000 | 0 |
+| FreiHAND/training/00032560 | 0 | 1 |
+| FreiHAND/training/00033560 | 1000 | 1 |
+| FreiHAND/training/00062560 | 30000 | 1 |
+| FreiHAND/training/00065120 | 0 | 2 |
+| FreiHAND/training/00066120 | 1000 | 2 |
+| FreiHAND/training/00095120 | 30000 | 2 |
+| FreiHAND/training/00097680 | 0 | 3 |
+| FreiHAND/training/00098680 | 1000 | 3 |
+| FreiHAND/training/00127680 | 30000 | 3 |
+
+
+## 全量自动验收
+
+4096 张上限的 33 个分片已完成转换及本地自动验收：130,240 张训练图片、3,960 张评估图片。逐批 RGB 码流/像素及 mask 数值核对通过，数组与规范登记、帧索引和样本唯一性检查通过。
+
+全量追加随机抽查已获用户确认；小样本报告与全量随机报告分开保留。运行日志仍只放 tmp。
+
+
+## 全量随机抽查
+
+[查看独立随机抽查报告](full_random/REPORT.md) · [打开 30 张随机样本预览](full_random/index.html)。全量随机抽查已获用户确认。
+
+
+发布确认：用户于 2026-09-18T17:01:03 确认全量随机抽查没有问题，允许发布。本报告仅描述实际完成的检查范围，不保证未抽中样本零缺陷。

FreiHAND/shared-v1.0/doc/validation/evidence.json

@@ -0,0 +1,90 @@
+{
+  "dataset": "FreiHAND",
+  "release": "shared-v1.0",
+  "schema_version": "hand-warehouse/1.0",
+  "scope": "15 selected images, not full release",
+  "human_review": "approved",
+  "full_release_validation": {
+    "automated": "passed",
+    "views": 33,
+    "counts": {
+      "training": 130240,
+      "evaluation": 3960
+    },
+    "completed_at_local": "2026-09-18T16:47:34",
+    "full_release_spot_review": "approved",
+    "random_review_ref": "full_random/samples.json",
+    "publication_approved_at_local": "2026-09-18T17:01:03"
+  },
+  "automated_checks_observed": [
+    "MANO_middle_MCP_root_recovery",
+    "all_mask_values_exact",
+    "all_rgb_packets_and_decoded_pixels_exact",
+    "tensor_serialization_exact"
+  ],
+  "preview_mask_vs_raw": "15 selected frames passed exact decoded-value comparison after native-bit-depth fix",
+  "scripts_sha256": {
+    "convert_freihand_lossless.py": "8beaece00bd4c34f38108d47fa470e0e353b2cefff3c4e5c49501ee4d720ac30",
+    "render_freihand_review.py": "fa989142aa946c8b1f93b52294d1e0485ab39024caa59bf3538f6e2a2f7f3831"
+  },
+  "reviewed_sample_snapshots": [
+    {
+      "sample_ids": [
+        "FreiHAND/evaluation/00000000",
+        "FreiHAND/evaluation/00001000",
+        "FreiHAND/evaluation/00003959"
+      ],
+      "converted_files_sha256": {
+        "annotation/keypoints/xyz.pt": "c43fb9a0e465dbe19da68bf391f7e1fa7e5ad25aa9cc08d4e57065ac02109387",
+        "annotation/keypoints/xyz_valid.pt": "79b7c71cbe3c99b5e0a3a158c1759e8e221af249bfd00eba3fd28d56ff273c7c",
+        "annotation/mano/motion.pt": "5f24b3792ff9270853665c6026e4da6580bf6c4fe3c6e3b58ca0af810eae4a3b",
+        "annotation/mano/motion_valid.pt": "088d70a1c9f4606e285e1765a39a86b47220bf86b6aaa877c5e58f4bd547da26",
+        "annotation/segment/instances.jsonl": "01ddb629790e3d5b06bfc1181226923ab232c1a4bd8365d50cb0e4a83369bae2",
+        "annotation/segment/mask.mkv": "345249f9f01f69a82e18b99815694896a14a6642f46f1844793fee65e8af54ba",
+        "annotation/segment/mask_valid.pt": "9360dc09feec98c40c5f6ae919161b1dfa2d6138797d859d2d76068ec5d597a8",
+        "camera/frames.jsonl": "b4c9c1e0a4100e17bf5154c873f155a60656615b8d10ff459f6da7a1560f168c",
+        "camera/intrinsic.pt": "cd1784d73dda9ba370ab8eb78d57363af0e5a7480949dcc1b38f5515c9d9613f",
+        "camera/intrinsic_valid.pt": "ef8ba0663fe4ca05b581ff385ccaa14e56e55317258d3f3e6997cbb2446cd721",
+        "metainfo.json": "8ecf14a2871fcd842ad7365d9146fa4e88f544c0b6aaa6b4c128f60b8ef7fa80",
+        "observation/rgb.mkv": "80480898983070cc67158ea2cb770247f0e70b8b6031fa94e55f303197fae34a",
+        "observation/rgb_valid.pt": "e3dead4bb3fe8f0370f944095c94a27e1443a83d690aa2e041611224aef2db28"
+      }
+    },
+    {
+      "sample_ids": [
+        "FreiHAND/training/00000000",
+        "FreiHAND/training/00001000",
+        "FreiHAND/training/00030000",
+        "FreiHAND/training/00032560",
+        "FreiHAND/training/00033560",
+        "FreiHAND/training/00062560",
+        "FreiHAND/training/00065120",
+        "FreiHAND/training/00066120",
+        "FreiHAND/training/00095120",
+        "FreiHAND/training/00097680",
+        "FreiHAND/training/00098680",
+        "FreiHAND/training/00127680"
+      ],
+      "converted_files_sha256": {
+        "annotation/keypoints/xyz.pt": "274d01ccfe77996c1713565450efc6d2e35b0284f03caefab08aa595da697e20",
+        "annotation/keypoints/xyz_valid.pt": "a45675ef37527c97db1753e961b6bc1dfe78ae733eb01417b2368ca180c16bc6",
+        "annotation/mano/motion.pt": "b0f94bf485472e87fc1d209c5c1238aa707ee9e61f1755a5411ab90f479be848",
+        "annotation/mano/motion_valid.pt": "6d029db1f1e3c5fdd0c53b231f0f4eec53691831ebc6b4c67c6169989e0a2134",
+        "annotation/segment/mask.mkv": "5337f40d49d1a180d0becd7870ad6d9783151e90b00054233e346d4b0b83ff34",
+        "annotation/segment/mask_valid.pt": "ea3aa6ed559bae7eed48ad8d27aebf6074a59d8b5a3cd46644fa8687d4b0dee6",
+        "camera/frames.jsonl": "c248c7a31013dbc46be29f898728df43756b284a7d0f39e0326d1bea569dc024",
+        "camera/intrinsic.pt": "87d79648af40e279e4a113e652e85d30b1bc3bc4b3b5421b4feb900ff42562e2",
+        "camera/intrinsic_valid.pt": "5f88f0322735f66aac5c0b83d092229b2ce6ec93a143e2882fd53ede459f6795",
+        "metainfo.json": "87eb58052b740bb7e388dd401f21c762af4c5fdc404cc440864981c14d6f39d5",
+        "observation/rgb.mkv": "39016875d319ab6809f093f5b1c50a5aa75c384ed1a2d59da5534bb5940c4eab",
+        "observation/rgb_valid.pt": "d025e58bc011843b707e210b32dc5d0854ee35ed617e08693c060f63f305993b"
+      }
+    }
+  ],
+  "human_review_confirmation": {
+    "reviewer": "user",
+    "confirmed_at_local": "2026-09-18T16:00:21",
+    "statement": "还是刚才那个审核页面吗？我看过了，没问题",
+    "scope": "15 previously displayed samples after preview bit-depth fix; approval to proceed to full conversion, not full-release validation"
+  }
+}