@jax-data-science/sci-image-visualizer 0.2.13 → 0.2.15

package/fesm2022/jax-data-science-sci-image-visualizer.mjs

@@ -425,6 +425,36 @@ const EYE_DEFAULTS = {
     overlapX: 0,
     overlapY: 0,
 };
+/**
+ * The retina is an elongated band spanning the whole field, so it crosses tile
+ * seams wherever they fall.
+ *
+ * With no overlap the tiles butt together at fixed boundaries: a structure
+ * sitting across a seam is cut in half, each tile sees only a slice, and the
+ * slice often fails to clear the confidence threshold — so the object vanishes
+ * exactly where the seam lands. Observed as multi-thousand-pixel holes in the
+ * middle of an otherwise continuous retina, which closed once overlap was
+ * raised.
+ *
+ * An object of size `s` is guaranteed to sit wholly inside some tile only when
+ * the overlap is at least `s / tileSize`; at 60% that covers anything up to
+ * ~307px of a 512px tile. The server's own table uses 0 here and has the same
+ * blind spot — this deliberately diverges from it.
+ */
+const RETINA_DEFAULTS = {
+    confidence: 0.6,
+    iouThreshold: 0.5,
+    // Cross-tile merging uses intersection-over-*smaller*, so at 0.3 a fragment is
+    // discarded once 30% of its own area is covered by a larger box — even when
+    // most of it lies outside. On a band detected as a chain of overlapping
+    // fragments that deletes the middle ones. Measured on one slide: 0.3 gave two
+    // boxes covering 75% of the band with a 3150px hole through it; 0.8 gave four
+    // boxes covering 100% with no hole, at the cost of two overlapping pairs.
+    // Matches the value jit-ui already sends the server for this checkpoint.
+    mergeThreshold: 0.8,
+    overlapX: 60,
+    overlapY: 60,
+};
 // Embryos are large, touch each other, and are imaged at native scale — hence
 // the heavy overlap and the far looser merge thresholds.
 const EMBRYO_DEFAULTS = {
@@ -445,7 +475,7 @@ const YOLO_MODELS = [
         id: 'yolov8x-seg-retina',
         label: 'Retina',
         modelUrl: `${HF}/yolov8x-seg-retina-onnx/resolve/main/model.fp16w.onnx`,
-        defaults: { ...EYE_DEFAULTS },
+        defaults: { ...RETINA_DEFAULTS },
     },
     {
         id: 'yolov8x-seg-embryo-m2',
@@ -1029,9 +1059,14 @@ class YoloDetectToolService {
             if (!ctx)
                 return null;
             ctx.drawImage(bitmap, 0, 0);
-            const out = ctx.getImageData(0, 0, bitmap.width, bitmap.height);
+            // Read the dimensions BEFORE close(): closing an ImageBitmap zeroes its
+            // width/height, so returning them afterwards yields a 0x0 image with a
+            // full pixel buffer — inference then runs on nothing and finds nothing.
+            const width = bitmap.width;
+            const height = bitmap.height;
+            const out = ctx.getImageData(0, 0, width, height);
             bitmap.close?.();
-            return { data: out.data, width: bitmap.width, height: bitmap.height };
+            return { data: out.data, width, height };
         }
         catch {
             return null;
@@ -8369,6 +8404,9 @@ class VisualizerComponent {
             mergeThreshold: def?.defaults.mergeThreshold ?? 0.3,
             overlapX: def?.defaults.overlapX ?? 0,
             overlapY: def?.defaults.overlapY ?? 0,
+            // Exposed in the dialog, so it must be seeded — otherwise the scale
+            // control reads as 0 and the re-crop never runs.
+            downsamplingFactor: 5,
             withMasks: true,
             simplifyTolerance: 1,
             minArea: 0,