fix(colmap_utils): handle 'images/'-prefixed keys in image_rename_map

nerfstudio/process_data/colmap_utils.py

@@ -450,7 +450,24 @@ def colmap_to_json(
 
         name = im_data.name
         if image_rename_map is not None:
-            name = image_rename_map[name]
+            # COLMAP stores im_data.name relative to the image_path it was
+            # given (e.g. "rig1/front/img.jpg"). The rename map is keyed by
+            # the source path relative to self.data, which can include an
+            # extra "images/" prefix when --data points at a parent
+            # containing an images/ subdirectory (common with --skip-colmap
+            # workflows). Try the bare name first, then the prefixed form.
+            # See https://github.com/nerfstudio-project/nerfstudio/issues/3759
+            if name in image_rename_map:
+                name = image_rename_map[name]
+            elif f"images/{name}" in image_rename_map:
+                name = image_rename_map[f"images/{name}"]
+            else:
+                raise KeyError(
+                    f"COLMAP image name {name!r} not in image_rename_map. "
+                    f"Tried {name!r} and 'images/' + {name!r}. "
+                    f"The map has {len(image_rename_map)} entries; "
+                    f"first 3 keys: {list(image_rename_map.keys())[:3]}."
+                )
         name = Path(f"./images/{name}")
 
         frame = {

tests/process_data/test_process_images.py

@@ -209,3 +209,75 @@ def test_process_images_recursively_skip_colmap(tmp_path: Path):
     )
     dataparser_poses = np.linalg.inv(dataparser_poses)
     np.testing.assert_allclose(original_poses, dataparser_poses, rtol=0, atol=1e-5)
+
+
+def test_process_images_data_parent_skip_colmap(tmp_path: Path):
+    """
+    Regression test for #3759: when --data points at a directory that contains
+    an `images/` subdirectory (a common layout for users running COLMAP externally
+    against `<data>/images/...` with --skip-colmap), the image_rename_map keys
+    were prefixed with `images/` while COLMAP's im_data.name was the bare path,
+    producing a KeyError in colmap_to_json. The fix accepts both key forms.
+    """
+    width = 100
+    height = 150
+    sparse_path = tmp_path / "sparse" / "0"
+    sparse_path.mkdir(exist_ok=True, parents=True)
+    # IMPORTANT: images live under <tmp_path>/images/ AND --data is <tmp_path>
+    # (the parent). This is the layout that triggered #3759.
+    (tmp_path / "images").mkdir(exist_ok=True, parents=True)
+    write_cameras_binary(
+        {1: Camera(1, "OPENCV", width, height, [110, 110, 50, 75, 0, 0, 0, 0, 0, 0])},
+        sparse_path / "cameras.bin",
+    )
+    write_points3D_binary(
+        {
+            1: Point3D(
+                id=1,
+                xyz=np.array([0, 0, 0]),
+                rgb=np.array([0, 0, 0]),
+                error=np.array([0]),
+                image_ids=np.array([1]),
+                point2D_idxs=np.array([0]),
+            ),
+        },
+        sparse_path / "points3D.bin",
+    )
+    frames = {}
+    num_frames = 5
+    qvecs = random_quaternion(num_frames)
+    tvecs = np.random.uniform(size=(num_frames, 3))
+    for i in range(num_frames):
+        # COLMAP stores the bare relative name (no "images/" prefix) because
+        # the user ran COLMAP with image_path=<tmp_path>/images.
+        frames[i + 1] = ColmapImage(i + 1, qvecs[i], tvecs[i], 1, f"image_{i}.png", [], [])
+        Image.new("RGB", (width, height)).save(tmp_path / "images" / f"image_{i}.png")
+    write_images_binary(frames, sparse_path / "images.bin")
+
+    # Mock missing COLMAP and ffmpeg in the dev env
+    old_path = os.environ.get("PATH", "")
+    os.environ["PATH"] = str(tmp_path / "mocked_bin") + f":{old_path}"
+    (tmp_path / "mocked_bin").mkdir()
+    (tmp_path / "mocked_bin" / "colmap").touch(mode=0o777)
+    (tmp_path / "mocked_bin" / "ffmpeg").touch(mode=0o777)
+
+    # Convert: --data points at <tmp_path> (the parent of images/), NOT at
+    # <tmp_path>/images/ like the existing tests do.
+    cmd = ImagesToNerfstudioDataset(
+        data=tmp_path, output_dir=tmp_path / "nerfstudio", colmap_model_path=sparse_path, skip_colmap=True
+    )
+    cmd.main()
+    os.environ["PATH"] = old_path
+
+    # Without the fix: KeyError before transforms.json is written.
+    assert (tmp_path / "nerfstudio" / "transforms.json").exists()
+    parser = NerfstudioDataParserConfig(
+        data=tmp_path / "nerfstudio",
+        downscale_factor=None,
+        orientation_method="none",
+        center_method="none",
+        auto_scale_poses=False,
+    ).setup()
+    outputs = parser.get_dataparser_outputs("train")
+    # All 5 frames should round-trip (allowing for the parser's train/eval split).
+    assert len(outputs.image_filenames) >= 1