Add SCAIL-Pose2 / SCAIL-2 workflow support

.gitignore

@@ -9,6 +9,10 @@ logs/
 .idea
 tools/
 .vscode/
+.venv/
+.sessions/
 convert_*
 *.pt
-*.pth
+*.pth
+.pla*/
+ref*/

MTV/nlf_bbox.py

@@ -0,0 +1,38 @@
+"""NLF bbox formatting helpers."""
+
+from __future__ import annotations
+
+
+def format_nlf_detected_boxes(all_boxes):
+    """Format detector xyxy rows for the public BBOX output."""
+
+    formatted_boxes = []
+    for box in all_boxes:
+        if hasattr(box, "detach"):
+            box = box.detach()
+        if hasattr(box, "cpu"):
+            box = box.cpu()
+        if hasattr(box, "numel") and box.numel() == 0:
+            formatted_boxes.append([0.0, 0.0, 0.0, 0.0])
+            continue
+        rows = box.tolist() if hasattr(box, "tolist") else box
+        if not rows:
+            formatted_boxes.append([0.0, 0.0, 0.0, 0.0])
+            continue
+        if isinstance(rows[0], (bool, int, float)):
+            rows = [rows]
+        candidates = []
+        for row in rows:
+            if len(row) < 4:
+                continue
+            x_min, y_min, x_max, y_max = (float(row[index]) for index in range(4))
+            if x_max <= x_min or y_max <= y_min:
+                continue
+            candidates.append([x_min, y_min, x_max, y_max])
+        if not candidates:
+            formatted_boxes.append([0.0, 0.0, 0.0, 0.0])
+        elif len(candidates) == 1:
+            formatted_boxes.append(candidates[0])
+        else:
+            formatted_boxes.append(candidates)
+    return formatted_boxes

MTV/nodes.py

@@ -15,6 +15,7 @@
 folder_paths.add_model_folder_path("nlf", os.path.join(folder_paths.models_dir, "nlf"))
 
 from .motion4d import SMPL_VQVAE, VectorQuantizer, Encoder, Decoder
+from .nlf_bbox import format_nlf_detected_boxes
 
 def check_jit_script_function():
     if torch.jit.script.__name__ != "script":
@@ -276,17 +277,9 @@ def predict(self, model, images, per_batch=-1):
             'joints3d_nonparam': [all_joints3d_nonparam],
         }
 
-        # Convert bboxes to list format: [x_min, y_min, x_max, y_max] for each detection
-        # Each box tensor is shape (1, 5) with [x_min, y_min, x_max, y_max, confidence]
-        formatted_boxes = []
-        for box in all_boxes:
-            # Handle empty detections (no person detected in frame)
-            if box.numel() == 0 or box.shape[0] == 0:
-                formatted_boxes.append([0.0, 0.0, 0.0, 0.0])
-            else:
-                # Extract first 4 values (x_min, y_min, x_max, y_max), drop confidence
-                bbox_values = box[0, :4].cpu().tolist()
-                formatted_boxes.append(bbox_values)
+        # Keep single-person frames backward compatible, but preserve all
+        # per-frame people for downstream multi-identity geometry matching.
+        formatted_boxes = format_nlf_detected_boxes(all_boxes)
 
         return (pose_results, formatted_boxes)