Mask tracker: keep only the cropped previous frame

step() cloned the whole RGBA frame to remember it (33 MB at 1080p,
133 MB at 4K) and then cropped that clone, although only the region
around the outline is ever read. The tracker now stores the previous
frame already cropped to the region the next step will use, computed
from the updated outline with the same margin and rounding as before,
so the crops, and therefore the tracking result, are unchanged.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

crates/track/src/mask.rs

@@ -44,6 +44,19 @@ pub fn bounds(p: &[[f64; 2]]) -> [f64; 4] {
     b
 }
 
+/// Crop `img` to `region` (layer pixels), with the layer offset of the crop's origin: the part
+/// of the frame the tracker reads. `None` when the region leaves too little of the frame.
+fn crop_region(img: &Image, off: [f64; 2], region: [f64; 4]) -> Option<(Image, [f64; 2])> {
+    let x0 = (region[0] + off[0]).floor().max(0.0) as i64;
+    let y0 = (region[1] + off[1]).floor().max(0.0) as i64;
+    let x1 = ((region[2] + off[0]).ceil() as i64).min(img.width as i64);
+    let y1 = ((region[3] + off[1]).ceil() as i64).min(img.height as i64);
+    if x1 <= x0 + 8 || y1 <= y0 + 8 {
+        return None;
+    }
+    Some((img.crop(x0, y0, (x1 - x0) as u32, (y1 - y0) as u32), [off[0] - x0 as f64, off[1] - y0 as f64]))
+}
+
 /// Mask tracker state.
 #[derive(Clone, Debug)]
 pub struct MaskTracker {
@@ -52,8 +65,8 @@ pub struct MaskTracker {
     pub outline: Vec<[f64; 2]>,
     /// The last frame-to-frame motion (a constant-velocity guess for the next frame).
     velocity: Homography,
-    prev: Option<Image>,
-    prev_offset: [f64; 2],
+    /// The previous frame cropped to the region the next step reads, with its layer offset.
+    prev: Option<(Image, [f64; 2])>,
     /// Features used on the last step (layer pixels), for display and tests.
     pub last_features: usize,
 }
@@ -72,44 +85,39 @@ pub struct MaskStep {
 impl MaskTracker {
     /// Start on `frame` with the mask `outline` (a closed polyline in layer pixels).
     pub fn new(model: Model, outline: Vec<[f64; 2]>, frame: &Frame) -> MaskTracker {
-        MaskTracker { model, outline, velocity: Homography::IDENTITY, prev: Some(frame.img.clone()), prev_offset: frame.offset, last_features: 0 }
+        let mut t = MaskTracker { model, outline, velocity: Homography::IDENTITY, prev: None, last_features: 0 };
+        t.prev = t.region().and_then(|r| crop_region(frame.img, frame.offset, r));
+        t
+    }
+
+    /// Crop region (layer pixels) of the next step: the mask plus room for its motion.
+    fn region(&self) -> Option<[f64; 4]> {
+        let b = bounds(&self.outline);
+        if !(b[0].is_finite() && b[2] > b[0] && b[3] > b[1]) {
+            return None;
+        }
+        let margin = (0.35 * (b[2] - b[0]).max(b[3] - b[1])).max(32.0);
+        Some([b[0] - margin, b[1] - margin, b[2] + margin, b[3] + margin])
     }
 
     /// Track into `frame` (the next frame in either direction). `None` when the mask's pixels
     /// can't be followed (no texture, or the mask left the frame).
     pub fn step(&mut self, frame: &Frame) -> Option<MaskStep> {
-        let prev = self.prev.take()?;
-        let res = self.step_inner(&prev, self.prev_offset, frame);
-        self.prev = Some(frame.img.clone());
-        self.prev_offset = frame.offset;
-        let r = res?;
-        self.outline = self.outline.iter().map(|p| r.motion.apply(*p)).collect();
-        self.velocity = r.motion;
-        Some(r)
+        let (ca, oa) = self.prev.take()?;
+        let res = self.step_inner(ca, oa, frame);
+        if let Some(r) = &res {
+            self.outline = self.outline.iter().map(|p| r.motion.apply(*p)).collect();
+            self.velocity = r.motion;
+        }
+        self.prev = self.region().and_then(|r| crop_region(frame.img, frame.offset, r));
+        res
     }
 
-    fn step_inner(&mut self, prev: &Image, prev_off: [f64; 2], frame: &Frame) -> Option<MaskStep> {
+    fn step_inner(&mut self, ca: Image, oa: [f64; 2], frame: &Frame) -> Option<MaskStep> {
         let b = bounds(&self.outline);
-        if !(b[0].is_finite() && b[2] > b[0] && b[3] > b[1]) {
-            return None;
-        }
         let (bw, bh) = (b[2] - b[0], b[3] - b[1]);
-        // Crop region (layer pixels): the mask plus room for its motion.
-        let margin = (0.35 * bw.max(bh)).max(32.0);
-        let (lx0, ly0, lx1, ly1) = (b[0] - margin, b[1] - margin, b[2] + margin, b[3] + margin);
-        // Both frames share one plane coordinate system: crop in the frames' image pixels.
-        let crop = |img: &Image, off: [f64; 2]| {
-            let x0 = (lx0 + off[0]).floor().max(0.0) as i64;
-            let y0 = (ly0 + off[1]).floor().max(0.0) as i64;
-            let x1 = ((lx1 + off[0]).ceil() as i64).min(img.width as i64);
-            let y1 = ((ly1 + off[1]).ceil() as i64).min(img.height as i64);
-            if x1 <= x0 + 8 || y1 <= y0 + 8 {
-                return None;
-            }
-            Some((img.crop(x0, y0, (x1 - x0) as u32, (y1 - y0) as u32), [off[0] - x0 as f64, off[1] - y0 as f64]))
-        };
-        let (ca, oa) = crop(prev, prev_off)?;
-        let (cb, ob) = crop(frame.img, frame.offset)?;
+        // Both frames share one plane coordinate system: the crops are in image pixels.
+        let (cb, ob) = crop_region(frame.img, frame.offset, self.region()?)?;
         let factor = analysis_factor(ca.width.max(cb.width), ca.height.max(cb.height), 720);
         // Pyramids must share coordinates: use a common origin (layer pixels → plane).
         let levels = 3;
@@ -219,4 +227,20 @@ mod tests {
             assert!(err < 1.5, "frame {f}: max vertex error {err}");
         }
     }
+
+    #[test]
+    fn follows_the_disk_through_a_frame_offset() {
+        let off = [40.0, -25.0];
+        let pose = |f: usize| Homography([[1.0, 0.0, 140.0 + 5.0 * f as f64], [0.0, 1.0, 110.0 - 3.0 * f as f64], [0.0, 0.0, 1.0]]);
+        let ring: Vec<[f64; 2]> = (0..24).map(|i| (i as f64 * 15.0).to_radians()).map(|a| [40.0 * a.cos(), 40.0 * a.sin()]).collect();
+        let layer = |f: usize| -> Vec<[f64; 2]> { ring.iter().map(|p| pose(f).apply(*p)).map(|q| [q[0] - off[0], q[1] - off[1]]).collect() };
+        let f0 = frame(&pose(0));
+        let mut mt = MaskTracker::new(Model::Translation, layer(0), &Frame { img: &f0, offset: off });
+        for f in 1..6 {
+            let img = frame(&pose(f));
+            mt.step(&Frame { img: &img, offset: off }).expect("tracked");
+            let err = mt.outline.iter().zip(&layer(f)).map(|(a, b)| (a[0] - b[0]).hypot(a[1] - b[1])).fold(0.0, f64::max);
+            assert!(err < 1.0, "frame {f}: max vertex error {err}");
+        }
+    }
 }