From c2c1fe5b8b07af161e9ce289cb3feacb4127254f Mon Sep 17 00:00:00 2001
From: Cameron Cordes <cameronc.dev@gmail.com>
Date: Thu, 30 Apr 2026 01:06:08 +0000
Subject: [PATCH] faces: bbox crop respects EXIF orientation + pads enough for
 RetinaFace
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two reasons manually-drawn bboxes were never resolving a face on
re-detection:

(1) The bbox arrives in display space (browser already applied EXIF
    orientation when rendering the carousel), but the `image` crate
    in crop_image_to_bbox opens raw pre-rotation pixels. For any
    phone photo with Orientation 6/8/etc., applying the bbox without
    rotating first crops a completely different region of the image
    — landing on background, hair, or empty pixels. Now reads the
    EXIF Orientation tag and applies it before indexing into the
    canonical-oriented dims.

(2) Padding was 10 % on each side. A typical 200×250 face bbox +
    10 % becomes ~240×300; insightface resizes that to det_size=640,
    so the face fills ~95 % of the input. RetinaFace's anchors
    expect faces at 20–60 % of input dimensions; at 95 % it
    routinely returns zero detections. Bumped to 50 % padding so the
    crop is 2× the bbox dims and the face occupies ~50 % of the
    input — anchor-friendly. Bbox is still clamped to image bounds,
    so edge-of-image cases just get less padding on the clipped
    side.

Together these explain why bbox-edit re-embed practically always
fell into the "no face detected" branch (and bbox-edit reverts
without the recent soft-fallback commit). Per-photo embedding
quality also improves slightly — same face, more context, better
landmarks for ArcFace.
---
 src/faces.rs | 30 ++++++++++++++++++++++++------
 1 file changed, 24 insertions(+), 6 deletions(-)

diff --git a/src/faces.rs b/src/faces.rs
index 6bcce3a..f45afa4 100644
--- a/src/faces.rs
+++ b/src/faces.rs
@@ -20,6 +20,7 @@
 
 use crate::Claims;
 use crate::ai::face_client::{DetectMeta, FaceClient, FaceDetectError};
+use crate::exif;
 use crate::database::schema::{face_detections, image_exif, persons};
 use crate::error::IntoHttpError;
 use crate::libraries::{self, Library};
@@ -2508,7 +2509,18 @@ fn crop_image_to_bbox(
     if nw <= 0.0 || nh <= 0.0 || nx + nw > 1.001 || ny + nh > 1.001 {
         return Err(anyhow!("bbox wh out of bounds or zero"));
     }
-    let img = image::open(abs_path).with_context(|| format!("open {:?}", abs_path))?;
+    let raw = image::open(abs_path).with_context(|| format!("open {:?}", abs_path))?;
+    // EXIF rotation: the bbox arrives in display space (the carousel /
+    // overlay are rendered post-rotation by the browser), but the
+    // `image` crate hands us raw pre-rotation pixels. For any phone
+    // photo with Orientation 6/8/etc., applying the bbox without
+    // rotating first lands the crop on a completely different region
+    // of the image — which is why manually-drawn bboxes basically
+    // never resolved a face on re-detection. Apply the orientation
+    // first, then index into the canonical-oriented dims. Photos with
+    // no EXIF rotation tag pay nothing (apply_orientation is a no-op).
+    let orientation = exif::read_orientation(abs_path).unwrap_or(1);
+    let img = exif::apply_orientation(raw, orientation);
     let (w, h) = img.dimensions();
     let px = (nx * w as f32).round().clamp(0.0, w as f32 - 1.0) as u32;
     let py = (ny * h as f32).round().clamp(0.0, h as f32 - 1.0) as u32;
@@ -2517,11 +2529,17 @@ fn crop_image_to_bbox(
     if pw == 0 || ph == 0 {
         return Err(anyhow!("crop produced zero-dim image"));
     }
-    // Pad the crop a bit so the detector has context — a tightly-drawn
-    // bbox often clips ears/jaw which hurts the embedding. 10% on each
-    // side is a reasonable default.
-    let pad_x = (pw / 10).max(1);
-    let pad_y = (ph / 10).max(1);
+    // Generous padding so RetinaFace has anchor-friendly context.
+    // Insightface internally resizes to det_size=640 (square). A
+    // tightly-drawn 200×250 face bbox + 10 % padding becomes ~240×300,
+    // which after resize fills ~95 % of the input — near the upper
+    // edge of RetinaFace's anchor scales, where it routinely returns
+    // zero detections. Padding to 50 % on each side makes the crop
+    // 2× the bbox dims (face occupies ~50 % of the input), where
+    // anchors hit cleanly. Bbox is clamped to image bounds, so
+    // edge-of-image bboxes just get less padding on the clipped side.
+    let pad_x = (pw / 2).max(1);
+    let pad_y = (ph / 2).max(1);
     let cx = px.saturating_sub(pad_x);
     let cy = py.saturating_sub(pad_y);
     let cw = (pw + 2 * pad_x).min(w - cx);