fix: simplify review — fix mask NMS bug, remove dead code, hoist imports

- _mask_nms: fix greedy suppression bug (j<=i compared raw indices,
not sort positions — masks escaped suppression). Use order[pos+1:]
- _mask_nms: batch GPU computation (single matmul + single CPU sync
instead of O(n^2) CUDA syncs per mask pair)
- _apply_nms: hoist torch import (already module-level) and
batched_nms import to module scope
- real-backend.js: remove dead isNotRelevant (always false after
mission_relevant filter)
- explain.js: clear _explainCache on job change to prevent stale
data across different videos

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

demo/js/explain.js

@@ -16,12 +16,19 @@ const LIGHTEN_MAP = {
 };
 
 let _explainAbort = null;
-const _explainCache = {};
+let _explainCache = {};
+let _explainCacheJobId = null;
 
 async function loadExplainability(jobId, trackId) {
   const panel = document.getElementById('explainPanel');
   if (!panel) return;
 
+  // Clear cache on job change
+  if (_explainCacheJobId !== jobId) {
+    _explainCache = {};
+    _explainCacheJobId = jobId;
+  }
+
   // Check cache
   if (_explainCache[trackId]) {
     renderExplainGraph(_explainCache[trackId], panel);

demo/js/real-backend.js

@@ -681,10 +681,9 @@ function renderTrackListFromData(tracks) {
 
   filtered.forEach((t, idx) => {
     const isSelected = ISR.STATE.selectedTrackId === t.track_id;
-    const isNotRelevant = t.mission_relevant === false;
 
     const card = document.createElement('div');
-    card.className = 'track-card' + (isSelected ? ' active' : '') + (isNotRelevant ? ' not-relevant' : '');
+    card.className = 'track-card' + (isSelected ? ' active' : '');
     card.dataset.trackId = t.track_id;
     card.style.animation = `track-card-enter 0.3s ease ${idx * 30}ms both`;
 

inference.py

@@ -316,6 +316,7 @@ def _build_detection_records(
 
 
 from utils.tracker import ByteTracker
+from utils.tiling import batched_nms as _batched_nms
 
 
 def _apply_nms(detections: List[Dict[str, Any]], iou_threshold: float = 0.45) -> List[Dict[str, Any]]:
@@ -330,13 +331,9 @@ def _apply_nms(detections: List[Dict[str, Any]], iou_threshold: float = 0.45) ->
     if len(detections) <= 1:
         return detections
 
-    import torch
-    from utils.tiling import batched_nms
-
     boxes = torch.tensor([d["bbox"] for d in detections], dtype=torch.float32)
     scores = torch.tensor([d["score"] for d in detections], dtype=torch.float32)
-    # Map labels to integer class IDs for batched NMS
-    label_to_id = {}
+    label_to_id: Dict[str, int] = {}
     label_ids = []
     for d in detections:
         lbl = d["label"]
@@ -345,7 +342,7 @@ def _apply_nms(detections: List[Dict[str, Any]], iou_threshold: float = 0.45) ->
         label_ids.append(label_to_id[lbl])
     labels = torch.tensor(label_ids, dtype=torch.int64)
 
-    keep = batched_nms(boxes, scores, labels, iou_threshold)
+    keep = _batched_nms(boxes, scores, labels, iou_threshold)
     return [detections[i] for i in keep.tolist()]
 
 

models/segmenters/grounded_sam2.py

@@ -95,28 +95,28 @@ class MaskDictionary:
         if n <= 1:
             return list(range(n))
 
-        # Compute mask areas
-        areas = [int(masks[i].sum()) for i in range(n)]
+        # Batch-compute areas and pairwise IoU on GPU (single sync)
+        flat = masks.view(n, -1).float()
+        areas = flat.sum(dim=1)  # (n,)
+        inter_matrix = flat @ flat.T  # (n, n)
+        union_matrix = areas.unsqueeze(1) + areas.unsqueeze(0) - inter_matrix
+        iou_matrix = (inter_matrix / union_matrix.clamp(min=1)).cpu().numpy()
+        areas_cpu = areas.cpu().numpy()
+
+        # Greedy suppression: sort by area descending, suppress smaller overlapping same-label masks
+        order = sorted(range(n), key=lambda i: areas_cpu[i], reverse=True)
         suppressed = [False] * n
-
-        # Sort by area descending (keep larger masks)
-        order = sorted(range(n), key=lambda i: areas[i], reverse=True)
-
         keep = []
-        for i in order:
+        for pos, i in enumerate(order):
             if suppressed[i]:
                 continue
             keep.append(i)
-            for j in order:
-                if j <= i or suppressed[j]:
+            for j in order[pos + 1:]:
+                if suppressed[j]:
                     continue
-                # Only suppress same-label masks
                 if labels[i] != labels[j]:
                     continue
-                # Compute mask IoU
-                inter = int((masks[i] & masks[j]).sum())
-                union = areas[i] + areas[j] - inter
-                if union > 0 and inter / union > iou_threshold:
+                if iou_matrix[i, j] > iou_threshold:
                     suppressed[j] = True
 
         return sorted(keep)