feat(RCLane): add CARLA dataset

dataset_carla.py

@@ -0,0 +1,83 @@
+"""
+CARLA dataset for RCLane (BanVienCorp/dataset_laneatt_fullmap, LaneATT format).
+
+Annotation is JSON-lines:
+  - first line: {"Ys": [72 y-anchors, 0..1065 step 15]}
+  - each record: {"lines": [[x per anchor, -2 = missing], ...],
+                  "types": [...], "image": "TownXX/.../frame.jpg"}
+Images are 1920x1080.
+
+Lanes are read in the original 1920x1080 space; the base class resizes to 800x320
+and runs `encode` to build the GT. RCLane generates its own GT, so no seg labels needed.
+"""
+
+import os
+import json
+
+import cv2
+
+from dataset import LaneEncodeDataset
+
+
+class CarlaLaneDataset(LaneEncodeDataset):
+    _MISSING = -2
+
+    def __init__(self, label_json, data_root, img_size=(320, 800),
+                 cache_dir=None, max_samples=None):
+        super().__init__(img_size, cache_dir)
+        self.data_root = data_root
+        with open(label_json) as f:
+            lines = f.read().splitlines()
+        self.Ys = json.loads(lines[0])["Ys"]
+        self.records = [r for r in (json.loads(ln) for ln in lines[1:])
+                        if "image" in r and "lines" in r]
+        if max_samples is not None:
+            self.records = self.records[:max_samples]
+
+    def __len__(self):
+        return len(self.records)
+
+    def _load(self, idx):
+        rec = self.records[idx]
+        img_path = os.path.join(self.data_root, rec["image"])
+        img = cv2.imread(img_path)
+        if img is None:
+            raise FileNotFoundError(img_path)
+        oh, ow = img.shape[:2]
+        lanes = []
+        for lane in rec["lines"]:
+            pts = [(x, self.Ys[i]) for i, x in enumerate(lane) if x != self._MISSING]
+            if len(pts) >= 2:
+                lanes.append(pts)
+        return img, lanes, ow, oh, rec["image"]
+
+
+# --------------------------------------------------------------------------- #
+#  smoke test -- builds a tiny FAKE CARLA sample (no real data needed) and
+#  verifies the loader parses annotations and produces GT tensors.
+# --------------------------------------------------------------------------- #
+if __name__ == "__main__":
+    import tempfile
+    import numpy as np
+
+    tmp = tempfile.mkdtemp(prefix="fake_carla_")
+    img_rel = "Town01/clear_noon/000000.jpg"
+    os.makedirs(os.path.join(tmp, os.path.dirname(img_rel)), exist_ok=True)
+    cv2.imwrite(os.path.join(tmp, img_rel), np.zeros((1080, 1920, 3), np.uint8))
+
+    ys = list(range(0, 1080, 15))                       # 72 anchors, like the real data
+    lane1 = [700 + i * 6 for i in range(len(ys))]       # x per anchor
+    lane2 = [1200 - i * 6 for i in range(len(ys))]
+    label = os.path.join(tmp, "label_train.json")
+    with open(label, "w") as f:
+        f.write(json.dumps({"Ys": ys}) + "\n")
+        f.write(json.dumps({"lines": [lane1, lane2], "types": [0, 0], "image": img_rel}) + "\n")
+
+    ds = CarlaLaneDataset(label_json=label, data_root=tmp, cache_dir=None)
+    print("dataset size:", len(ds))
+    x, tgt = ds[0]
+    print("image:", tuple(x.shape), "| seg fg pixels:", int((tgt["seg_map"] > 0).sum()))
+    for k, v in tgt.items():
+        assert tuple(v.shape)[-2:] == (320, 800), f"bad shape for {k}"
+    assert int((tgt["seg_map"] > 0).sum()) > 0, "no foreground -- parse/encode failed"
+    print("OK -- CARLA loader parses JSONL and produces GT.")