增加数据集

.gitignore

@@ -102,4 +102,31 @@ CCPD_GREEN_TRANS/
 ccpd_combined_5k5k_spe/images/
 ccpd_combined_5k5k_spe/labels/
 output/
-build/
+
+# ============================================================
+# Local dataset payloads (keep README only)
+# ============================================================
+CCPD_322K_SMALLEST_5000_640x352/images/
+CCPD_322K_SMALLEST_5000_640x352/labels/
+CCPD_322K_SMALLEST_5000_640x352/metadata/
+CCPD_322K_SMALLEST_5000_640x352/dataset.yaml
+CCPD_RK3568_SMALL_HARD_V1/images/
+CCPD_RK3568_SMALL_HARD_V1/labels/
+CCPD_RK3568_SMALL_HARD_V1/metadata/
+CCPD_RK3568_SMALL_HARD_V1/dataset.yaml
+
+# Archived/generated payloads; retain README files only.
+archive/**/images/
+archive/**/labels/
+archive/**/logs/
+archive/**/metadata/
+archive/**/dataset.yaml
+archive/**/test_*.txt
+
+# Local sample images at the dataset repository root.
+/*.jpg
+/*.jpeg
+
+# Local one-off dataset generator.
+/scripts/build_ccpd_diverse_30k.py
+build/

archive/CCPD_DIVERSE_30K_WITH_OCCLUSION_20260814/README.md

@@ -0,0 +1,15 @@
+# CCPD_DIVERSE_30K
+
+本目录由 `scripts/build_ccpd_diverse_30k.py` 生成，完全使用本地数据，不联网。
+
+- 总目标：30,000 张
+- 类型比例：blue 35%、green 35%、other 30%
+- 输入尺寸：640×352 与 352×640 约各半
+- 尺寸指标：`sqrt(车牌框面积)`，8 个尺寸区间
+- 难度：simple / hard 约各半
+- 训练集：`images/train` 与 `labels/train`，混合所有类型
+- 测试集：`images/test/blue|green|other`，按类型分类
+- 元数据：`metadata/manifest.jsonl`、`metadata/dataset_stats.json`
+- 生成日志：`logs/generation.log`
+
+同一 CCPD 源图按 `source_id` 只分配到一个 split；由同源图生成的变体不会跨 train/test。