use yaml file when using --data flag

Author: mikel-brostrom

boxmot/engine/cli.py

@@ -61,6 +61,50 @@ def resolve_dataset_cfg_path(data_ref) -> Optional[Path]:
     return None
 
 
+def resolve_source_from_data_ref(data_ref: str, overwrite: bool = False) -> tuple[str, str, str]:
+    """
+    Resolve a data/source reference into (source_path, benchmark_name, split_name).
+
+    Supports either direct dataset paths or dataset yaml references.
+    """
+    source_path = Path(str(data_ref))
+    benchmark = source_path.parent.name
+    split = source_path.name
+
+    cfg_path = resolve_dataset_cfg_path(data_ref)
+    if cfg_path is None:
+        return str(data_ref), benchmark, split
+
+    cfg = load_dataset_cfg_path(cfg_path)
+    bench_name = Path(cfg["benchmark"]["source"]).name
+    dataset_url = cfg["download"]["dataset_url"]
+
+    if dataset_url:
+        if dataset_url.startswith("hf://"):
+            dataset_dest = TRACKEVAL / bench_name
+        else:
+            dataset_dest = TRACKEVAL / f"{bench_name}.zip"
+    else:
+        dataset_dest = Path(cfg["download"].get("dataset_dest", f"assets/{bench_name}"))
+
+    download_eval_data(
+        runs_url=cfg["download"]["runs_url"],
+        dataset_url=dataset_url,
+        dataset_dest=dataset_dest,
+        overwrite=overwrite,
+    )
+
+    split = cfg["benchmark"]["split"]
+    if dataset_url:
+        resolved_source = TRACKEVAL / f"{bench_name}/{split}"
+    elif "source" in cfg["benchmark"]:
+        resolved_source = Path(cfg["benchmark"]["source"]) / split
+    else:
+        resolved_source = dataset_dest / split
+
+    return str(resolved_source), bench_name, split
+
+
 def ensure_model_extension(model_path):
     """
     Ensure model path has .pt extension.
@@ -376,8 +420,7 @@ def track(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwarg
     if tracker:
         kwargs['tracking_method'] = tracker
     src = kwargs.pop('source')
-    source_path = Path(src)
-    bench, split = source_path.parent.name, source_path.name
+    resolved_source, bench, split = resolve_source_from_data_ref(src)
 
     # Auto-append .pt extension if missing
     yolo_model = ensure_model_extension(yolo_model)
@@ -387,37 +430,10 @@ def track(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwarg
               'yolo_model': yolo_model,
               'reid_model': reid_model,
               'classes': parse_classes(classes),
-              'source': src,
+              'source': resolved_source,
               'benchmark': bench,
               'split': split}
     args = SimpleNamespace(**params)
-    
-    # 2) if doing MOT17/20-ablation, pull down the dataset and rewire args.source/split
-    cfg_path = resolve_dataset_cfg_path(args.source)
-    if cfg_path is not None:
-        cfg = load_dataset_cfg_path(cfg_path)
-        
-        # Determine dataset destination
-        if cfg["download"]["dataset_url"]:
-            dataset_dest = TRACKEVAL / f"{Path(cfg['benchmark']['source']).name}.zip"
-        else:
-            # For custom datasets without URL, use the path from config if available, or default to assets
-            dataset_dest = Path(cfg["download"].get("dataset_dest", f"assets/{Path(cfg['benchmark']['source']).name}"))
-
-        download_eval_data(
-            runs_url=cfg["download"]["runs_url"],
-            dataset_url=cfg["download"]["dataset_url"],
-            dataset_dest=dataset_dest,
-            overwrite=False
-        )
-        args.benchmark = Path(cfg["benchmark"]["source"]).name
-        args.split = cfg["benchmark"]["split"]
-        if cfg["download"]["dataset_url"]:
-            args.source = TRACKEVAL / f"{args.benchmark}/{args.split}"
-        elif "source" in cfg["benchmark"]:
-            args.source = Path(cfg["benchmark"]["source"]) / args.split
-        else:
-            args.source = dataset_dest / args.split
 
     from boxmot.engine.tracker import main as run_track
     run_track(args)
@@ -437,8 +453,7 @@ def generate(ctx, detector, reid, yolo_model, reid_model, classes, **kwargs):
     if reid:
         reid_model = [ensure_model_extension(reid)]
     src = kwargs.pop('data')
-    source_path = Path(src)
-    bench, split = source_path.parent.name, source_path.name
+    resolved_source, bench, split = resolve_source_from_data_ref(src)
 
     # Auto-append .pt extension if missing
     yolo_model = [ensure_model_extension(m) for m in yolo_model]
@@ -449,7 +464,7 @@ def generate(ctx, detector, reid, yolo_model, reid_model, classes, **kwargs):
               'reid_model': list(reid_model),
               'classes': parse_classes(classes),
               'data': src,
-              'source': src,
+              'source': resolved_source,
               'benchmark': bench,
               'split': split}
     args = SimpleNamespace(**params)
@@ -475,8 +490,6 @@ def eval(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwargs
     if tracker:
         kwargs['tracking_method'] = tracker
     src = kwargs.pop('data')
-    source_path = Path(src)
-    bench, split = source_path.parent.name, source_path.name
 
     # Auto-append .pt extension if missing
     yolo_model = [ensure_model_extension(m) for m in yolo_model]
@@ -487,9 +500,6 @@ def eval(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwargs
               'reid_model': list(reid_model),
               'classes': parse_classes(classes),
               'data': src,
-              'source': src,
-              'benchmark': bench,
-              'split': split,
               'imgsz': [1088, 1920]}
     args = SimpleNamespace(**params)
     from boxmot.engine.evaluator import main as run_eval
@@ -515,8 +525,6 @@ def tune(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwargs
     if tracker:
         kwargs['tracking_method'] = tracker
     src = kwargs.pop('data')
-    source_path = Path(src)
-    bench, split = source_path.parent.name, source_path.name
 
     # Auto-append .pt extension if missing
     yolo_model = [ensure_model_extension(m) for m in yolo_model]
@@ -526,10 +534,7 @@ def tune(ctx, detector, reid, tracker, yolo_model, reid_model, classes, **kwargs
               'yolo_model': list(yolo_model),
               'reid_model': list(reid_model),
               'classes': parse_classes(classes),
-              'data': src,
-              'source': src,
-              'benchmark': bench,
-              'split': split}
+              'data': src}
     args = SimpleNamespace(**params)
     from boxmot.engine.tuner import main as run_tuning
     run_tuning(args)

boxmot/engine/evaluator.py

@@ -24,15 +24,15 @@
 from boxmot.utils.plots import MetricsPlotter
 from boxmot.utils.misc import increment_path, prompt_overwrite
 from boxmot.utils.timing import TimingStats, wrap_tracker_reid
-from typing import Optional, List, Dict, Generator, Union
+from typing import Optional, List, Dict, Generator
 
 from boxmot.utils.dataloaders.MOT17 import MOT17DetEmbDataset
 from boxmot.postprocessing.gsi import gsi
 
 from boxmot.engine.inference import DetectorReIDPipeline, extract_detections, filter_detections
 from boxmot.detectors import default_imgsz
 from boxmot.utils.mot_utils import convert_to_mot_format, write_mot_results
-from boxmot.utils.download import download_eval_data, download_trackeval
+from boxmot.utils.download import download_trackeval
 
 checker = RequirementsChecker()
 checker.check_packages(('ultralytics', ))  # install
@@ -57,47 +57,6 @@ def load_dataset_cfg(name: str) -> dict:
         return yaml.safe_load(f)
 
 
-def load_dataset_cfg_path(path: Path) -> dict:
-    """Load a dataset config directly from a yaml file path."""
-    with open(path, 'r') as f:
-        return yaml.safe_load(f)
-
-
-def resolve_dataset_cfg_path(data_ref: Union[str, Path, None]) -> Optional[Path]:
-    """
-    Resolve dataset config references from --data/--source.
-
-    Supports:
-    - config name: "MOT17-ablation"
-    - config filename: "MOT17-ablation.yaml"
-    - explicit path: "/path/to/custom.yaml" or "./custom.yaml"
-    """
-    if data_ref is None:
-        return None
-
-    ref = Path(str(data_ref))
-
-    # Explicit local yaml path (absolute or relative)
-    if ref.suffix in {".yaml", ".yml"} and ref.exists() and ref.is_file():
-        return ref.resolve()
-
-    # Config filename in boxmot/configs/datasets
-    if ref.suffix in {".yaml", ".yml"}:
-        cfg_by_filename = DATASET_CONFIGS / ref.name
-        if cfg_by_filename.exists() and cfg_by_filename.is_file():
-            return cfg_by_filename
-        cfg_by_stem = DATASET_CONFIGS / f"{ref.stem}.yaml"
-        if cfg_by_stem.exists() and cfg_by_stem.is_file():
-            return cfg_by_stem
-
-    # Config name in boxmot/configs/datasets
-    cfg_by_name = DATASET_CONFIGS / f"{str(data_ref)}.yaml"
-    if cfg_by_name.exists() and cfg_by_name.is_file():
-        return cfg_by_name
-
-    return None
-
-
 def eval_init(args,
               trackeval_dest: Path = TRACKEVAL,
               branch: str = "master",
@@ -111,33 +70,28 @@ def eval_init(args,
     # 1) download the TrackEval code
     download_trackeval(dest=trackeval_dest, branch=branch, overwrite=overwrite)
 
-    # 2) if using a dataset yaml, pull down data (if needed) and rewire args.source/split
-    data_ref = getattr(args, "data", None) or getattr(args, "source", None)
-    cfg_path = resolve_dataset_cfg_path(data_ref)
-    if cfg_path is not None:
-        cfg = load_dataset_cfg_path(cfg_path)
-        
-        # Determine dataset destination
-        if cfg["download"]["dataset_url"]:
-            dataset_dest = TRACKEVAL / f"{Path(cfg['benchmark']['source']).name}.zip"
-        else:
-            # For custom datasets without URL, use the path from config if available, or default to assets
-            dataset_dest = Path(cfg["download"].get("dataset_dest", f"assets/{Path(cfg['benchmark']['source']).name}"))
-
-        download_eval_data(
-            runs_url=cfg["download"]["runs_url"],
-            dataset_url=cfg["download"]["dataset_url"],
-            dataset_dest=dataset_dest,
-            overwrite=overwrite
-        )
-        args.benchmark = Path(cfg["benchmark"]["source"]).name
-        args.split = cfg["benchmark"]["split"]
-        if cfg["download"]["dataset_url"]:
-            args.source = TRACKEVAL / f"{args.benchmark}/{args.split}"
-        elif "source" in cfg["benchmark"]:
-            args.source = Path(cfg["benchmark"]["source"]) / args.split
+    # 2) resolve dataset reference via shared CLI helper.
+    # For eval/tune, prefer --data and do not use --source when data is provided.
+    from boxmot.engine.cli import resolve_source_from_data_ref
+
+    data_ref = getattr(args, "data", None)
+    if data_ref is not None:
+        resolved_source, benchmark, split = resolve_source_from_data_ref(data_ref, overwrite=overwrite)
+        args.source = resolved_source
+        args.benchmark = benchmark
+        args.split = split
+    else:
+        # Backward-compatible fallback for callers that pass source directly.
+        source_path = Path(str(getattr(args, "source", "")))
+        if source_path.exists() and source_path.is_dir():
+            args.source = source_path
+            args.benchmark = getattr(args, "benchmark", source_path.parent.name)
+            args.split = getattr(args, "split", source_path.name)
         else:
-            args.source = dataset_dest / args.split
+            resolved_source, benchmark, split = resolve_source_from_data_ref(getattr(args, "source", None), overwrite=overwrite)
+            args.source = resolved_source
+            args.benchmark = benchmark
+            args.split = split
 
     # 3) finally, make source an absolute Path everywhere
     args.source = Path(args.source).resolve()
@@ -1123,6 +1077,8 @@ def run_trackeval(opt: argparse.Namespace, verbose: bool = True) -> dict:
 
 
 def main(args):
+    data_ref = getattr(args, "data", getattr(args, "source", None))
+
     # Print evaluation pipeline header (blue palette)
     LOGGER.info("")
     LOGGER.opt(colors=True).info("<blue>" + "="*60 + "</blue>")
@@ -1131,7 +1087,7 @@ def main(args):
     LOGGER.opt(colors=True).info(f"<bold>Detector:</bold>  <cyan>{args.yolo_model[0]}</cyan>")
     LOGGER.opt(colors=True).info(f"<bold>ReID:</bold>      <cyan>{args.reid_model[0]}</cyan>")
     LOGGER.opt(colors=True).info(f"<bold>Tracker:</bold>   <cyan>{args.tracking_method}</cyan>")
-    LOGGER.opt(colors=True).info(f"<bold>Benchmark:</bold> <cyan>{args.source}</cyan>")
+    LOGGER.opt(colors=True).info(f"<bold>Benchmark:</bold> <cyan>{data_ref}</cyan>")
     LOGGER.opt(colors=True).info(f"<bold>Image size:</bold> <cyan>{getattr(args, 'imgsz', None)}</cyan>")
     LOGGER.opt(colors=True).info("<blue>" + "="*60 + "</blue>")