Add spherical_reprojected mode to panorama_sfm example

benchmark/reconstruction/evaluate.py

@@ -65,6 +65,7 @@
 from evaluation.tartanair import (
     DatasetTartanAirPerspective,
     DatasetTartanAirSpherical,
+    DatasetTartanAirSphericalReprojected,
 )
 from evaluation.tartanair.tartanair_v2 import load_manifest
 from evaluation.utils import (
@@ -92,6 +93,7 @@ def run_once(args: argparse.Namespace) -> MetricsByDatasetByCatByScene | None:
         "imc2025": DatasetIMC2025,
         "tartanair-v2-perspective": DatasetTartanAirPerspective,
         "tartanair-v2-spherical": DatasetTartanAirSpherical,
+        "tartanair-v2-spherical-reprojected": DatasetTartanAirSphericalReprojected,
     }
 
     metrics: MetricsByDatasetByCatByScene = {}
@@ -190,6 +192,9 @@ def run_seeds(args: argparse.Namespace) -> None:
 
 
 def main() -> None:
+    import multiprocessing
+
+    multiprocessing.set_start_method("spawn", force=True)
     args = parse_args(__doc__)
     if args.seeds is not None:
         run_seeds(args)

benchmark/reconstruction/evaluation/tartanair/README.md

@@ -30,12 +30,13 @@ this subset:
 The exact source revision, trajectories, frames, and packaging checksums are
 recorded in `tartanair_v2_manifest.json`.
 
-From the COLMAP checkout, download and evaluate both pipeline variants with:
+From the COLMAP checkout, download and evaluate the pipeline variants with:
 
 ```bash
 python benchmark/reconstruction/download.py --datasets tartanair-v2
 python benchmark/reconstruction/evaluate.py \
   --datasets tartanair-v2-perspective tartanair-v2-spherical \
+    tartanair-v2-spherical-reprojected \
   --colmap_path /path/to/colmap
 ```
 

benchmark/reconstruction/evaluation/tartanair/__init__.py

@@ -199,3 +199,8 @@ class DatasetTartanAirPerspective(DatasetTartanAir):
 class DatasetTartanAirSpherical(DatasetTartanAir):
     dataset_name = "tartanair-v2-spherical"
     render_type = "spherical"
+
+
+class DatasetTartanAirSphericalReprojected(DatasetTartanAir):
+    dataset_name = "tartanair-v2-spherical-reprojected"
+    render_type = "spherical_reprojected"

benchmark/reconstruction/evaluation/utils.py

@@ -790,8 +790,19 @@ def panorama_reconstruction(
     if sparse_path.exists():
         pycolmap.logging.info("Skipping reconstruction, as it already exists")
         return
-    if args.feature != "sift":
-        raise ValueError("Panorama reconstruction currently supports SIFT only")
+    if args.feature == "sift":
+        extractor_type = pycolmap.FeatureExtractorType.SIFT
+        matcher_type = pycolmap.FeatureMatcherType.SIFT_BRUTEFORCE
+    elif args.feature == "aliked":
+        # ALIKED_LIGHTGLUE rather than automatic_reconstructor's plain
+        # "--feature aliked" default (ALIKED_BRUTEFORCE): this benchmark
+        # exists to evaluate the ALIKED+LightGlue combination specifically.
+        extractor_type = pycolmap.FeatureExtractorType.ALIKED_N16ROT
+        matcher_type = pycolmap.FeatureMatcherType.ALIKED_LIGHTGLUE
+    else:
+        raise ValueError(
+            f"Unknown feature type for panorama reconstruction: {args.feature}"
+        )
     if args.mapper == "hierarchical":
         raise ValueError(
             "Panorama reconstruction does not support hierarchical mapping"
@@ -802,7 +813,11 @@ def panorama_reconstruction(
         )
 
     render_type = scene_info.reconstruction_backend.removeprefix("panorama-")
-    if render_type not in {"perspective_overlapping", "spherical"}:
+    if render_type not in {
+        "perspective_overlapping",
+        "spherical",
+        "spherical_reprojected",
+    }:
         raise ValueError(
             f"Unknown panorama reconstruction backend: "
             f"{scene_info.reconstruction_backend}"
@@ -831,6 +846,8 @@ def panorama_reconstruction(
             matcher=panorama.Matcher.SEQUENTIAL,
             mapper=panorama.Mapper(args.mapper),
             render_type=panorama.PanoRenderType(render_type),
+            extractor_type=extractor_type,
+            matcher_type=matcher_type,
             random_seed=args.random_seed,
             num_threads=num_threads,
             gpu_index=gpu_index,