[feat] evaluate multisensor odometry in CI

cuvslam-skills/cuvslam-ci/reference.md

@@ -68,11 +68,13 @@ Repository secrets, split read from write so fork-reachable jobs never hold a ke
 ## Dataset registry and layout
 
 - `DATASETS` in `tools/python_tools/cuvslam_tools/dataset_registry.py` is the single source of truth. A `DatasetSpec` holds the ID, the preparation module, and its `EvalSpec` records; an `EvalSpec` holds the reporter config filename, the `cuvslam_app` flags, suite membership, and gating. Provisionable means the dataset is present; eval-enabled means it has at least one `EvalSpec`. KITTI, EuRoC, TUM, ICL-NUIM and M3ED-SPOT are eval-enabled; `tartan` and `coda` are provisionable only. Smoke runs KITTI, EuRoC and ICL-NUIM, which covers stereo, stereo-inertial and RGB-D; TUM is full-only because it is the larger RGB-D corpus and ICL-NUIM already covers the modality pre-merge, and M3ED-SPOT is full-only because at 56 GiB and 57k frames in two modes it is the most expensive record in the suite.
+- One dataset can carry several records. KITTI, TUM and ICL-NUIM each have a second, full-only `informational` record that replays the same reporter config in `multisensor` mode, so the `MSF` KPI rows compare the cuNLS solver against the default one on identical frames. EuRoC and M3ED-SPOT have none: the cuNLS solver projects through a pinhole camera and only warns on the fisheye and polynomial models those two use, so a record there would report numbers computed from the wrong geometry. `multisensor` needs `USE_CUNLS=ON`, which every CI configuration has.
 - The module is standard library only and imports converters lazily, so shell wrappers call it with `PYTHONPATH=tools/python_tools` inside `cuvslam-ci:local` before anything is installed. `datasets_config.sh` wraps it as `dataset_registry`; `run_eval.sh` defines its own shim because the S3 variables `datasets_config.sh` requires are absent in the eval container.
 - Subcommands: `validate [--dataset] [--suite]`, `list [--eval] [--suite]`, `eval-records [--suite]` (tab-separated `id`, KPI prefix, config path, flags), `kpi-keys [--suite]`, `prepare-module`, `prepare --root-file`, `verify-staged --root`.
 - Suites: `EVAL_SUITE` selects records in `stage_eval_datasets.sh`, `check_eval_prerequisites.sh`, and `run_eval.sh`, and `eval_cuvslam_in_docker.sh` forwards it into the container. Unset means every record; validation requires every `EvalSpec` to belong to `full`, so unset and `full` agree. `validate --suite` additionally rejects a suite that would select nothing.
 - `kpi-keys --suite` prints the `<PREFIX>_<METRIC>_<TYPE>_<MODE>` keys a suite can produce, so a consumer can tell a key that is legitimately absent from one that went missing. It lists both `ODOM` and `SLAM` for every record because which of the two a run emits depends on `sequence_title` inside the reporter config, which ships in the tarball. `KPI_METRICS` and `ODOMETRY_MODE_TYPES` in the registry mirror `REQUIRED_METRICS` and `odometry_mode_to_type` in `scripts/cuvslam_kpi_report.py` and have to be edited together; compare them with `python3 -m cuvslam_tools.dataset_registry kpi-keys --suite full` against the keys in `scripts/kpi_baseline_ranges.json` after changing either side.
 - Derived, never declared: `<id>.tar`, the staged directory, the `/sequences` mount, and the KPI prefix (first hyphen-delimited token of the config filename, upper-cased). Validation rejects two records that derive the same prefix and `--odometry_mode`.
+- The reporter writes to `$CUVSLAM_OUTPUT/<config stem>-<odometry mode>/<timestamp>/`. The mode is part of the directory because the KPI collector reads only the newest run under each directory, so two records sharing one config would otherwise overwrite each other; the prefix is unaffected because it is the first hyphen-delimited token. KPI types are `MCAM`, `MONO`, `VIO`, `RGBD` and `MSF`, one per odometry mode, and an unrecognized mode is rejected rather than filed under another mode's keys.
 - Tarball: uncompressed `<id>.tar` at `<S3_DATASETS_BUCKET>/<id>.tar`, whose root is the directory `prepare()` returned. Staged to `<RUNNER_LOCAL_DATASETS_ROOT>/datasets/vslam/<id>/` and mounted read-only into the eval container at `/datasets`. An ETag file skips re-download when the cache is current. After extraction, `verify-staged` checks the shipped config's `dataset_folder` equals `<id>/`.
 
 ## KPI outputs

scripts/cuvslam_kpi_report.py

@@ -40,7 +40,7 @@
 
 
 def display_dataset_key(key):
-    """TARTAN_FLAKY-STEREO_ODOM -> TARTAN_F-STEREO_ODOM (display only)."""
+    """TARTAN_FLAKY-MCAM_ODOM -> TARTAN_F-MCAM_ODOM (display only)."""
     for full, short in DATASET_DISPLAY_ALIASES.items():
         if key.startswith(full + "-"):
             return short + key[len(full):]
@@ -78,28 +78,43 @@ def parse_all_stats_json(json_path):
         return None
 
 
+# One KPI type per odometry mode. Keep in step with ODOMETRY_MODE_TYPES in
+# tools/python_tools/cuvslam_tools/dataset_registry.py, which names the keys a
+# suite is expected to produce: a mismatch reads as a missing KPI. A type must
+# not contain an underscore, because parse_kpi_key splits keys on it.
+ODOMETRY_MODE_TYPES = {
+    'multicamera': 'MCAM',
+    'mono': 'MONO',
+    'inertial': 'VIO',
+    'rgbd': 'RGBD',
+    'multisensor': 'MSF',
+}
+
+
 def odometry_mode_to_type(odometry_mode):
     """Convert odometry_mode string to dataset type.
 
     Args:
-        odometry_mode: String like "OdometryMode.Multicamera". Matching is
-            case-insensitive, so command-line values like "multicamera" map the
-            same way. None or non-string values fall back to STEREO.
+        odometry_mode: String like "OdometryMode.Multicamera". The enum qualifier
+            is optional and matching is case-insensitive, so command-line values
+            like "multicamera" map the same way.
 
     Returns:
-        str: Dataset type (MONO, STEREO, VIO, RGBD)
+        str: Dataset type (MCAM, MONO, VIO, RGBD, MSF)
+
+    Raises:
+        ValueError: If the mode is unrecognized. Defaulting would file the run
+            under another mode's KPI keys and overwrite them.
     """
-    normalized = str(odometry_mode).lower()
-    if 'multicamera' in normalized:
-        return 'STEREO'
-    elif 'mono' in normalized:
-        return 'MONO'
-    elif 'inertial' in normalized:
-        return 'VIO'
-    elif 'rgbd' in normalized:
-        return 'RGBD'
-    else:
-        return 'STEREO'
+    # Exact lookup on the unqualified name, not substring containment: a value
+    # such as "notmultisensor" has to stay unmapped so the caller skips the run
+    # instead of recording it as MSF.
+    normalized = str(odometry_mode).rsplit('.', 1)[-1].lower()
+    if normalized in ODOMETRY_MODE_TYPES:
+        return ODOMETRY_MODE_TYPES[normalized]
+    raise ValueError(
+        f"unknown odometry_mode {odometry_mode!r}; expected one of {', '.join(ODOMETRY_MODE_TYPES)}"
+    )
 
 
 def load_baseline_ranges(path):
@@ -210,20 +225,19 @@ def process_dataset_folder(dataset_folder_path):
         print(f'Warning: failed to parse all_stats.json in {stats_folder}')
         return None
 
-    if all_stats and 'odometry_mode' in all_stats[0]:
+    # The mode is the only trustworthy source of the KPI type. Guessing it from
+    # the folder name mislabels every config whose name mentions another mode,
+    # such as kitti-vio_slam_gt run as multicamera.
+    if 'odometry_mode' not in all_stats[0]:
+        print(f'Warning: odometry_mode not found in {all_stats_json}; cannot name this run\'s KPI keys')
+        return None
+
+    try:
         dataset_type = odometry_mode_to_type(all_stats[0]['odometry_mode'])
-        print(f'  Detected dataset type: {dataset_type} (from odometry_mode: {all_stats[0]["odometry_mode"]})')
-    else:
-        print(f'  Warning: odometry_mode not found in JSON, falling back to folder name parsing')
-        folder_name = os.path.basename(dataset_folder_path).lower()
-        if 'mono' in folder_name:
-            dataset_type = 'MONO'
-        elif 'vio' in folder_name or 'imu' in folder_name:
-            dataset_type = 'VIO'
-        elif 'rgbd' in folder_name or 'depth' in folder_name:
-            dataset_type = 'RGBD'
-        else:
-            dataset_type = 'STEREO'
+    except ValueError as exc:
+        print(f'Warning: {exc} in {all_stats_json}')
+        return None
+    print(f'  Detected dataset type: {dataset_type} (from odometry_mode: {all_stats[0]["odometry_mode"]})')
 
     odom_stats = [s for s in all_stats if 'ODOM' in s.get('sequence_title', '').upper()]
     slam_stats = [s for s in all_stats if 'SLAM' in s.get('sequence_title', '').upper()]

scripts/kpi_baseline_ranges.json

@@ -12,16 +12,16 @@
         "trusted nightly kpi_<date>.json into 'expected' below and pick a tolerance."
     ],
     "kpis": {
-        "KITTI_ATE_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "KITTI_ARE_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "KITTI_Kabsch_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "KITTI_TrackingLosts_STEREO_ODOM": {"expected": null, "tol_abs": 1},
-        "KITTI_FPS_STEREO_ODOM": {"expected": null, "tol_pct": 15},
-        "KITTI_ATE_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "KITTI_ARE_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "KITTI_Kabsch_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "KITTI_TrackingLosts_STEREO_SLAM": {"expected": null, "tol_abs": 1},
-        "KITTI_FPS_STEREO_SLAM": {"expected": null, "tol_pct": 15},
+        "KITTI_ATE_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_ARE_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_Kabsch_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_TrackingLosts_MCAM_ODOM": {"expected": null, "tol_abs": 1},
+        "KITTI_FPS_MCAM_ODOM": {"expected": null, "tol_pct": 15},
+        "KITTI_ATE_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_ARE_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_Kabsch_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_TrackingLosts_MCAM_SLAM": {"expected": null, "tol_abs": 1},
+        "KITTI_FPS_MCAM_SLAM": {"expected": null, "tol_pct": 15},
         "EUROC_ATE_VIO_ODOM": {"expected": null, "tol_pct": 10},
         "EUROC_ARE_VIO_ODOM": {"expected": null, "tol_pct": 10},
         "EUROC_Kabsch_VIO_ODOM": {"expected": null, "tol_pct": 10},
@@ -52,15 +52,45 @@
         "ICL_NUIM_Kabsch_RGBD_SLAM": {"expected": null, "tol_pct": 10},
         "ICL_NUIM_TrackingLosts_RGBD_SLAM": {"expected": null, "tol_abs": 1},
         "ICL_NUIM_FPS_RGBD_SLAM": {"expected": null, "tol_pct": 15},
-        "M3ED_SPOT_ATE_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_ARE_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_Kabsch_STEREO_ODOM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_TrackingLosts_STEREO_ODOM": {"expected": null, "tol_abs": 1},
-        "M3ED_SPOT_FPS_STEREO_ODOM": {"expected": null, "tol_pct": 15},
-        "M3ED_SPOT_ATE_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_ARE_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_Kabsch_STEREO_SLAM": {"expected": null, "tol_pct": 10},
-        "M3ED_SPOT_TrackingLosts_STEREO_SLAM": {"expected": null, "tol_abs": 1},
-        "M3ED_SPOT_FPS_STEREO_SLAM": {"expected": null, "tol_pct": 15}
+        "M3ED_SPOT_ATE_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_ARE_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_Kabsch_MCAM_ODOM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_TrackingLosts_MCAM_ODOM": {"expected": null, "tol_abs": 1},
+        "M3ED_SPOT_FPS_MCAM_ODOM": {"expected": null, "tol_pct": 15},
+        "M3ED_SPOT_ATE_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_ARE_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_Kabsch_MCAM_SLAM": {"expected": null, "tol_pct": 10},
+        "M3ED_SPOT_TrackingLosts_MCAM_SLAM": {"expected": null, "tol_abs": 1},
+        "M3ED_SPOT_FPS_MCAM_SLAM": {"expected": null, "tol_pct": 15},
+        "KITTI_ATE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_ARE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_Kabsch_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "KITTI_TrackingLosts_MSF_ODOM": {"expected": null, "tol_abs": 1},
+        "KITTI_FPS_MSF_ODOM": {"expected": null, "tol_pct": 15},
+        "KITTI_ATE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_ARE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_Kabsch_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "KITTI_TrackingLosts_MSF_SLAM": {"expected": null, "tol_abs": 1},
+        "KITTI_FPS_MSF_SLAM": {"expected": null, "tol_pct": 15},
+        "TUM_ATE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "TUM_ARE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "TUM_Kabsch_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "TUM_TrackingLosts_MSF_ODOM": {"expected": null, "tol_abs": 1},
+        "TUM_FPS_MSF_ODOM": {"expected": null, "tol_pct": 15},
+        "TUM_ATE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "TUM_ARE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "TUM_Kabsch_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "TUM_TrackingLosts_MSF_SLAM": {"expected": null, "tol_abs": 1},
+        "TUM_FPS_MSF_SLAM": {"expected": null, "tol_pct": 15},
+        "ICL_NUIM_ATE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_ARE_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_Kabsch_MSF_ODOM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_TrackingLosts_MSF_ODOM": {"expected": null, "tol_abs": 1},
+        "ICL_NUIM_FPS_MSF_ODOM": {"expected": null, "tol_pct": 15},
+        "ICL_NUIM_ATE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_ARE_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_Kabsch_MSF_SLAM": {"expected": null, "tol_pct": 10},
+        "ICL_NUIM_TrackingLosts_MSF_SLAM": {"expected": null, "tol_abs": 1},
+        "ICL_NUIM_FPS_MSF_SLAM": {"expected": null, "tol_pct": 15}
     }
 }

tools/python_tools/cuvslam_tools/dataset_registry.py

@@ -61,16 +61,17 @@
 KPI_METRICS = ("ATE", "ARE", "Kabsch", "TrackingLosts", "FPS")
 KPI_MODES = ("ODOM", "SLAM")
 ODOMETRY_MODE_TYPES = {
-    "multicamera": "STEREO",
+    "multicamera": "MCAM",
     "mono": "MONO",
     "inertial": "VIO",
     "rgbd": "RGBD",
+    "multisensor": "MSF",
 }
 
-# Values cuvslam_app accepts for --odometry_mode. cuvslam_kpi_report.py maps each
-# to a distinct KPI type, and anything unrecognized there falls back to STEREO,
-# so an unknown mode here would silently mislabel a KPI key.
-ODOMETRY_MODES = ("multicamera", "mono", "inertial", "rgbd")
+# Values cuvslam_app accepts for --odometry_mode, which are exactly the modes
+# cuvslam_kpi_report.py can name a KPI type for. It rejects anything else rather
+# than filing the run under another mode's keys.
+ODOMETRY_MODES = tuple(ODOMETRY_MODE_TYPES)
 
 _DATASET_ID = re.compile(r"^[a-z0-9_]+$")
 _KPI_PREFIX = re.compile(r"^[A-Z0-9_]+$")
@@ -111,8 +112,18 @@ def odometry_mode(self) -> str:
 
     @property
     def kpi_type(self) -> str:
-        """Dataset type the KPI collector will derive from the odometry mode."""
-        return ODOMETRY_MODE_TYPES.get(self.odometry_mode, "STEREO")
+        """Dataset type the KPI collector will derive from the odometry mode.
+
+        Raises:
+            RegistryError: when the mode has no type, which ``validate`` also
+                rejects. Falling back to one would name another mode's keys.
+        """
+        try:
+            return ODOMETRY_MODE_TYPES[self.odometry_mode]
+        except KeyError:
+            raise RegistryError(
+                f"eval '{self.config}': no KPI type for odometry mode {self.odometry_mode!r}"
+            ) from None
 
     def kpi_keys(self) -> tuple[str, ...]:
         """KPI keys this record can produce, as `<PREFIX>_<METRIC>_<TYPE>_<MODE>`.
@@ -176,6 +187,16 @@ def _rgbd_args() -> tuple[str, ...]:
     return ("--odometry_mode=rgbd", "--async_sba=false", "--use_segments")
 
 
+def _multisensor_args(*extra: str) -> tuple[str, ...]:
+    """Flags for the unified multi-sensor mode.
+
+    Only registered for rigs the cuNLS solver can model. It projects through a
+    pinhole camera and merely warns on any other model, so EuRoC (fisheye) and
+    M3ED-SPOT (polynomial) would report numbers computed from the wrong geometry.
+    """
+    return ("--odometry_mode=multisensor", *extra, "--async_sba=false", "--use_segments")
+
+
 DATASETS: dict[str, DatasetSpec] = {
     "kitti": DatasetSpec(
         dataset_id="kitti",
@@ -186,6 +207,16 @@ def _rgbd_args() -> tuple[str, ...]:
                 args=_stereo_args("--rectified_stereo_camera=true"),
                 suites=frozenset(SUITES),
             ),
+            # Same config, same frames, same ground truth as the record above, so
+            # the two KPI rows compare the cuNLS solver against the default one
+            # directly. The rig carries no depth; multi-sensor qualifies here on
+            # the overlapping stereo pair alone.
+            EvalSpec(
+                config="kitti-vio_slam_gt.cfg",
+                args=_multisensor_args("--rectified_stereo_camera=true", "--multicam_mode=moderate"),
+                suites=frozenset({FULL_SUITE}),
+                gating="informational",
+            ),
         ),
     ),
     "euroc": DatasetSpec(
@@ -216,6 +247,12 @@ def _rgbd_args() -> tuple[str, ...]:
                 args=_rgbd_args(),
                 suites=frozenset({FULL_SUITE}),
             ),
+            EvalSpec(
+                config="tum-rgbd_slam.cfg",
+                args=_multisensor_args(),
+                suites=frozenset({FULL_SUITE}),
+                gating="informational",
+            ),
         ),
     ),
     # One config in both suites, not one per suite: a second would derive its own
@@ -229,6 +266,14 @@ def _rgbd_args() -> tuple[str, ...]:
                 args=_rgbd_args(),
                 suites=frozenset(SUITES),
             ),
+            # Full only while the mode is experimental, even though the dataset
+            # is cheap enough for smoke. Promote once the MSF keys are calibrated.
+            EvalSpec(
+                config="icl_nuim-rgbd_slam.cfg",
+                args=_multisensor_args(),
+                suites=frozenset({FULL_SUITE}),
+                gating="informational",
+            ),
         ),
     ),
     # Full only, and by a wide margin the most expensive record: 56 GiB to stage
@@ -312,7 +357,7 @@ def _validate_eval(spec: DatasetSpec, record: EvalSpec) -> None:
     if modes[0] not in ODOMETRY_MODES:
         raise RegistryError(
             f"{where}: --odometry_mode must be one of {', '.join(ODOMETRY_MODES)}; "
-            "an unrecognized mode silently becomes a STEREO KPI type"
+            "the KPI collector has no type for anything else"
         )
     if not record.suites:
         raise RegistryError(f"{where}: must belong to at least one suite")