Klaus 280 discard pose data when convex hull is larger than expected mouse size

main.nf

@@ -40,7 +40,7 @@ workflow{
         // Integrate annotations back into pose files
         // This branch requires files to be local and already url-ified
         PREPARE_DATA(params.input_batch, params.location, true)
-        INTEGRATE_CORNER_ANNOTATIONS(PREPATE_DATA.out.file_processing_channel, params.sleap_file)
+        INTEGRATE_CORNER_ANNOTATIONS(PREPARE_DATA.out.file_processing_channel, params.sleap_file)
         ADD_DUMMY_VIDEO(INTEGRATE_CORNER_ANNOTATIONS.out, params.clip_duration)
         paired_video_and_pose = ADD_DUMMY_VIDEO.out[0]
 
@@ -50,7 +50,7 @@ workflow{
     if (params.workflow == "single-mouse-v6-features"){
         PREPARE_DATA(params.input_batch, params.location, false)
         // Generate features from pose_v6 files
-        ADD_DUMMY_VIDEO(PREPARE_DATA.out.out_file, params.clip_duration)
+        ADD_DUMMY_VIDEO(PREPARE_DATA.out.file_processing_channel, params.clip_duration)
         paired_video_and_pose = ADD_DUMMY_VIDEO.out[0]
         SINGLE_MOUSE_V6_FEATURES(paired_video_and_pose)
     }

nextflow/configs/profiles/sumner2.config

@@ -192,16 +192,16 @@ process {
      * Runtime options
      */
     withLabel: "tracking" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/deployment-runtime_2025-08-26.sif"
+        container = "/projects/kumar-lab/meta/images/mouse-tracking-runtime/runtime/v0.1.2/latest.sif"
     }
     withLabel: "jabs_classify" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/JABS-GUI_2025-02-12_v0.18.1.sif"
+        container = "/projects/kumar-lab/meta/images/JABS-behavior-classifier/headless/v0.36.1/latest.sif"
     }
     withLabel: "jabs_postprocess" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/JABS-Postprocessing-2025-03-27_864d687.sif"
+        container = "/projects/kumar-lab/meta/images/JABS-postprocess/jabs-postprocess/v0.3.1/latest.sif"
     }
     withLabel: "jabs_table_convert" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/support-r-code_2025-02-11.sif"
+        container = "/projects/kumar-lab/meta/images/mouse-tracking-runtime/RBase/v0.1.2/latest.sif"
     }
     withLabel: "gait" {
         container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/gait-pipeline-2025-03-27.sif"
@@ -210,10 +210,10 @@ process {
         container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/vfi-2025-03-27.sif"
     }
     withLabel: "sleap" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/sleap-1.4.1.sif"
+        container = "/projects/kumar-lab/meta/images/mouse-tracking-runtime/sleap-1.4.1/v0.1.2/latest.sif"
     }
     withLabel: "sleap_io" {
-        container = "/projects/kumar-lab/multimouse-pipeline/nextflow-containers/sleap-io-0.2.0.sif"
+        container = "/projects/kumar-lab/meta/images/mouse-tracking-runtime/sleap-io-0.2.0/v0.1.2/latest.sif"
     }
     withLabel: "rclone" {
         // executor.queueSize = 1

nextflow/workflows/io.nf

@@ -54,9 +54,9 @@ workflow PREPARE_DATA {
 
     // Files should be appropriately URLified to avoid collisions within the pipeline
     if (skip_urlify) {
-        file_processing_channel = file_batch.readLines().flatMap { line -> file(line) }
+        file_processing_channel = file_batch.splitText().map { line -> file(line.trim()) }
     } else {
-        file_processing_channel = URLIFY_FILE(file_batch.readLines().flatMap(), params.path_depth).file
+        file_processing_channel = URLIFY_FILE(file_batch.splitText().map { it.trim() }, params.path_depth).file
     }
 
     emit:

pyproject.toml

@@ -28,6 +28,7 @@ gpu = [
   "torch==2.6.0",
   "torchvision==0.21.0",
   "torchaudio==2.6.0",
+  "nvidia-cusparselt-cu12==0.6.3",
 ]
 
 # CPU-only convenience for local tests (unchanged idea)

src/mouse_tracking/cli/utils.py

@@ -6,13 +6,15 @@
 from rich import print
 
 from mouse_tracking import __version__
+from mouse_tracking.core.config.pose_utils import PoseUtilsConfig
 from mouse_tracking.matching.match_predictions import match_predictions
 from mouse_tracking.pose import render
-from mouse_tracking.pose.convert import downgrade_pose_file
 from mouse_tracking.utils import fecal_boli, static_objects
 from mouse_tracking.utils.clip_video import clip_video_auto, clip_video_manual
+from mouse_tracking.utils.writers import downgrade_pose_file, filter_large_poses
 
 app = typer.Typer()
+CONFIG = PoseUtilsConfig()
 
 
 def version_callback(value: bool) -> None:
@@ -248,3 +250,22 @@ def stitch_tracklets(
     This command stitches tracklets from the specified source.
     """
     match_predictions(in_pose)
+
+
+@app.command()
+def filter_large_area_pose(
+    in_pose: Path = typer.Argument(..., help="Input HDF5 pose file"),
+    max_area: int = typer.Option(
+        CONFIG.OFA_MAX_EXPECTED_AREA_PX,
+        help="Maximum area a pose can have, using a bounding box on keypoint pose.",
+    ),
+):
+    """
+    Filer pose by area.
+
+    This command unmarks identity of pose with large areas.
+    """
+    filter_large_poses(
+        in_pose,
+        max_area,
+    )

src/mouse_tracking/core/config/pose_utils.py

@@ -36,4 +36,4 @@ class PoseUtilsConfig(BaseSettings):
     MIN_JABS_KEYPOINTS: int = 3
 
     # Large animals are rarely larger than 100px in our OFA
-    OFA_MAX_EXPECTED_AREA_PX = 150 * 150
+    OFA_MAX_EXPECTED_AREA_PX: int = 150 * 150

src/mouse_tracking/pose/convert.py

@@ -1,14 +1,8 @@
 """Pose data conversion utilities."""
 
-import os
-import re
-
-import h5py
 import numpy as np
 
-from mouse_tracking.core.exceptions import InvalidPoseFileException
 from mouse_tracking.utils.run_length_encode import run_length_encode
-from mouse_tracking.utils.writers import write_pixel_per_cm_attr, write_pose_v2_data
 
 
 def v2_to_v3(pose_data, conf_data, threshold: float = 0.3):
@@ -95,53 +89,3 @@ def multi_to_v2(pose_data, conf_data, identity_data):
         return_list.append((cur_id, single_pose, single_conf))
 
     return return_list
-
-
-def downgrade_pose_file(pose_h5_path, disable_id: bool = False):
-    """Downgrades a multi-mouse pose file into multiple single mouse pose files.
-
-    Args:
-            pose_h5_path: input pose file
-            disable_id: bool to disable identity embedding tracks (if available) and use tracklet data instead
-    """
-    if not os.path.isfile(pose_h5_path):
-        raise FileNotFoundError(f"ERROR: missing file: {pose_h5_path}")
-    # Read in all the necessary data
-    with h5py.File(pose_h5_path, "r") as pose_h5:
-        if "version" in pose_h5["poseest"].attrs:
-            major_version = pose_h5["poseest"].attrs["version"][0]
-        else:
-            raise InvalidPoseFileException(
-                f"Pose file {pose_h5_path} did not have a valid version."
-            )
-        if major_version == 2:
-            print(f"Pose file {pose_h5_path} is already v2. Exiting.")
-            exit(0)
-
-        all_points = pose_h5["poseest/points"][:]
-        all_confidence = pose_h5["poseest/confidence"][:]
-        if major_version >= 4 and not disable_id:
-            all_track_id = pose_h5["poseest/instance_embed_id"][:]
-        elif major_version >= 3:
-            all_track_id = pose_h5["poseest/instance_track_id"][:]
-        try:
-            config_str = pose_h5["poseest/points"].attrs["config"]
-            model_str = pose_h5["poseest/points"].attrs["model"]
-        except (KeyError, AttributeError):
-            config_str = "unknown"
-            model_str = "unknown"
-        pose_attrs = pose_h5["poseest"].attrs
-        if "cm_per_pixel" in pose_attrs and "cm_per_pixel_source" in pose_attrs:
-            pixel_scaling = True
-            px_per_cm = pose_h5["poseest"].attrs["cm_per_pixel"]
-            source = pose_h5["poseest"].attrs["cm_per_pixel_source"]
-        else:
-            pixel_scaling = False
-
-    downgraded_pose_data = multi_to_v2(all_points, all_confidence, all_track_id)
-    new_file_base = re.sub("_pose_est_v[0-9]+\\.h5", "", pose_h5_path)
-    for animal_id, pose_data, conf_data in downgraded_pose_data:
-        out_fname = f"{new_file_base}_animal_{animal_id}_pose_est_v2.h5"
-        write_pose_v2_data(out_fname, pose_data, conf_data, config_str, model_str)
-        if pixel_scaling:
-            write_pixel_per_cm_attr(out_fname, px_per_cm, source)

src/mouse_tracking/utils/writers.py

@@ -1,13 +1,16 @@
 """Functions related to saving data to pose files."""
 
+import os
+import re
 from pathlib import Path
 
 import h5py
 import numpy as np
 
 from mouse_tracking.core.exceptions import InvalidPoseFileException
 from mouse_tracking.matching import hungarian_match_points_seg
-from mouse_tracking.pose.convert import v2_to_v3
+from mouse_tracking.pose.convert import multi_to_v2, v2_to_v3
+from mouse_tracking.pose.inspect import get_pose_bounding_box
 
 
 def promote_pose_data(pose_file, current_version: int, new_version: int):
@@ -588,3 +591,92 @@ def write_pose_clip(
         for key, attrs in all_attrs.items():
             for cur_attr, data in attrs.items():
                 out_f[key].attrs.create(cur_attr, data)
+
+
+def downgrade_pose_file(pose_h5_path, disable_id: bool = False):
+    """Downgrades a multi-mouse pose file into multiple single mouse pose files.
+
+    Args:
+            pose_h5_path: input pose file
+            disable_id: bool to disable identity embedding tracks (if available) and use tracklet data instead
+    """
+    if not os.path.isfile(pose_h5_path):
+        raise FileNotFoundError(f"ERROR: missing file: {pose_h5_path}")
+    # Read in all the necessary data
+    with h5py.File(pose_h5_path, "r") as pose_h5:
+        if "version" in pose_h5["poseest"].attrs:
+            major_version = pose_h5["poseest"].attrs["version"][0]
+        else:
+            raise InvalidPoseFileException(
+                f"Pose file {pose_h5_path} did not have a valid version."
+            )
+        if major_version == 2:
+            print(f"Pose file {pose_h5_path} is already v2. Exiting.")
+            exit(0)
+
+        all_points = pose_h5["poseest/points"][:]
+        all_confidence = pose_h5["poseest/confidence"][:]
+        if major_version >= 4 and not disable_id:
+            all_track_id = pose_h5["poseest/instance_embed_id"][:]
+        elif major_version >= 3:
+            all_track_id = pose_h5["poseest/instance_track_id"][:]
+        try:
+            config_str = pose_h5["poseest/points"].attrs["config"]
+            model_str = pose_h5["poseest/points"].attrs["model"]
+        except (KeyError, AttributeError):
+            config_str = "unknown"
+            model_str = "unknown"
+        pose_attrs = pose_h5["poseest"].attrs
+        if "cm_per_pixel" in pose_attrs and "cm_per_pixel_source" in pose_attrs:
+            pixel_scaling = True
+            px_per_cm = pose_h5["poseest"].attrs["cm_per_pixel"]
+            source = pose_h5["poseest"].attrs["cm_per_pixel_source"]
+        else:
+            pixel_scaling = False
+
+    downgraded_pose_data = multi_to_v2(all_points, all_confidence, all_track_id)
+    new_file_base = re.sub("_pose_est_v[0-9]+\\.h5", "", pose_h5_path)
+    for animal_id, pose_data, conf_data in downgraded_pose_data:
+        out_fname = f"{new_file_base}_animal_{animal_id}_pose_est_v2.h5"
+        write_pose_v2_data(out_fname, pose_data, conf_data, config_str, model_str)
+        if pixel_scaling:
+            write_pixel_per_cm_attr(out_fname, px_per_cm, source)
+
+
+def filter_large_poses(in_pose_f: str | Path, area_threshold: float):
+    """Unmarks identity of poses that exceed area threshold.
+
+    Args:
+        in_pose_f: Input pose filename
+        area_threshold: maximum pose bounding box allowed
+
+    Raises:
+        InvalidPoseFileException if the pose file is not >= 4.
+    """
+    with h5py.File(in_pose_f, "r") as f:
+        try:
+            current_version = f["poseest"].attrs["version"][0]
+        except (KeyError, AttributeError, IndexError):
+            InvalidPoseFileException("Pose file does not have a version.")
+        if current_version < 4:
+            raise InvalidPoseFileException(
+                f"Pose file {in_pose_f} is {current_version}. Filtering is only implemented for pose file versions > 4."
+            )
+
+        pose_data = f["poseest/points"][:]
+        pose_confidence = f["poseest/confidence"][:]
+        identity_data = f["poseest/instance_embed_id"][:]
+        pose_masks = f["poseest/id_mask"][:]
+
+    pose_boxes = get_pose_bounding_box(pose_data, pose_confidence)
+    pose_boxes = pose_boxes.astype(float)
+    pose_box_size = pose_boxes[:, :, 1] - pose_boxes[:, :, 0]
+    pose_box_area = pose_box_size[:, :, 0] * pose_box_size[:, :, 1]
+
+    identities_to_unassign = np.where(pose_box_area > area_threshold)
+    identity_data[identities_to_unassign] = 0
+    pose_masks[identities_to_unassign] = 1
+
+    with h5py.File(in_pose_f, "a") as f:
+        f["poseest/instance_embed_id"][:] = identity_data
+        f["poseest/id_mask"][:] = pose_masks

vm/rclone.def

@@ -0,0 +1,6 @@
+from: rclone/rclone
+bootstrap: docker
+
+%post
+    apk add --no-cache bash
+