Class BenchmarkDataset

benchmark/datasets/benchmark_dataset.py:104–374

Deterministic scene-level benchmark dataset. Each __getitem__ returns all fixed frames of one scene, used for model inference and evaluation. No randomness, no data augmentation. Data layout (SpatialBenchmark, pre-split by view_density): / / /<sc

class BenchmarkDataset(torch.utils.data.Dataset):
    """Deterministic scene-level benchmark dataset.

    Each __getitem__ returns all fixed frames of one scene, used for model inference and evaluation.
    No randomness, no data augmentation.

    Data layout (SpatialBenchmark, pre-split by view_density):
        <benchmark_root>/<density>/<dataset>/<scene_path>/{images,depths,poses,intrinsics,...,meta.json}
    Where density in {single, sparse, medium, dense}; each scene folder&#x27;s images/ is already filtered
    per frame, so use list(range(n_frames)) as positional indices to load all frames of the scene.

    Args:
        scene_index_path: path to scene_index.json (e.g. all_scenes.json)
        benchmark_root: SpatialBenchmark root directory
        tags: tag filter (e.g. ["sparse", "indoor"])
        tag_operator: "AND" or "OR"
        max_scenes: maximum number of scenes (used for quick testing)
        conf_threshold: Ropedia confidence threshold
    """

    # ImageNet normalization
    IMG_NORM = T.Compose([
        T.ToTensor(),
        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])

    # SpatialBenchmark root pre-split by density (per-frame data is read from here):
    #   <root>/<density>/<dataset>/<scene_path>/{images,depths,...,meta.json}
    DEFAULT_BENCHMARK_ROOT = "SpatialBenchmark"

    def __init__(self,
                 scene_index_path,
                 benchmark_root=None,
                 tags=None,
                 tag_operator="AND",
                 max_scenes=None,
                 resolution_override=None,
                 conf_threshold=0.3,
                 shuffle_seed=None,
                 priority_datasets=None):

        self.registry = TagRegistry(scene_index_path)
        self.shuffle_seed = shuffle_seed

        # Filter scenes
        if tags:
            self.scenes = self.registry.query(tags, operator=tag_operator)
        else:
            self.scenes = self.registry.scenes

        if max_scenes:
            self.scenes = self.scenes[:max_scenes]

        # Move high-VRAM / OOM-prone datasets to the front of the queue, so OOM is triggered on the
        # first scene; combined with run_dense_benchmark.py's OOM detection this kills the process
        # immediately and skips to the next model
        if priority_datasets:
            priority_set = set(priority_datasets)