immich-automated-selfie-tim.../src/pipeline/steps/alignment.rs
2026-02-12 20:20:12 +01:00

340 lines
11 KiB
Rust

//! Eye-based face alignment step.
//!
//! Aligns faces based on eye positions to ensure consistent eye placement
//! across all images in the timelapse.
//!
//! Uses a single affine transformation matrix that combines rotation, scaling,
//! and translation to align eyes at the desired positions in one operation.
use crate::config::Config;
use crate::pipeline::{computed_keys, PipelineContext, Point, ProcessingStep, StepOutcome};
use async_trait::async_trait;
use image::{DynamicImage, Rgb, RgbImage};
/// Aligns faces based on eye positions.
///
/// This step:
/// 1. Retrieves landmarks from ctx.computed["landmarks"]
/// 2. Calculates a single affine transformation matrix that:
/// - Rotates to make eyes horizontal
/// - Scales to match desired inter-eye distance
/// - Translates to position eyes at configured target positions
/// 3. Applies the transformation using bilinear interpolation
pub struct AlignmentStep;
#[async_trait]
impl ProcessingStep for AlignmentStep {
fn id(&self) -> &'static str {
"alignment"
}
fn name(&self) -> &'static str {
"Alignment"
}
async fn execute(&self, mut ctx: PipelineContext, config: &Config) -> StepOutcome {
// Get landmarks from previous step (required)
let landmarks: crate::pipeline::Landmarks = match ctx
.get_computed(computed_keys::LANDMARKS)
.and_then(|v| v.as_landmarks())
{
Some(l) => l.clone(),
None => {
// Landmarks should always be available since LandmarksStep is mandatory
return StepOutcome::Error {
ctx,
error: "Landmarks not available for alignment - pipeline misconfigured"
.to_string(),
};
}
};
let image = match ctx.take_image("alignment") {
Ok(img) => img,
Err(e) => return StepOutcome::Error { ctx, error: e },
};
let output_size = config.processing.output.size;
// Get eye centers
let left_eye = landmarks.left_eye_center();
let right_eye = landmarks.right_eye_center();
// Guard against coincident eyes (would cause division by zero in transform)
let eye_distance =
((right_eye.x - left_eye.x).powi(2) + (right_eye.y - left_eye.y).powi(2)).sqrt();
if eye_distance < 1e-6 {
return StepOutcome::Skip {
ctx,
reason: "alignment".to_string(),
detail: Some("Eye positions are too close together for alignment".to_string()),
};
}
// Calculate the transformation matrix
let transform = calculate_eye_alignment_transform(
left_eye,
right_eye,
output_size,
&config.processing.alignment,
);
// Apply the transformation
let aligned = apply_affine_transform(&image, &transform, output_size);
ctx.image = Some(aligned);
StepOutcome::Continue(ctx)
}
}
/// 2x3 affine transformation matrix.
/// Represents the transformation: [x', y'] = [[a, b, c], [d, e, f]] * [x, y, 1]
#[derive(Debug, Clone, Copy)]
struct AffineMatrix {
a: f32,
b: f32,
c: f32,
d: f32,
e: f32,
f: f32,
}
impl AffineMatrix {
/// Create a translation matrix.
fn translation(tx: f32, ty: f32) -> Self {
Self {
a: 1.0,
b: 0.0,
c: tx,
d: 0.0,
e: 1.0,
f: ty,
}
}
/// Create a rotation matrix (angle in radians).
fn rotation(angle: f32) -> Self {
let cos_a = angle.cos();
let sin_a = angle.sin();
Self {
a: cos_a,
b: -sin_a,
c: 0.0,
d: sin_a,
e: cos_a,
f: 0.0,
}
}
/// Create a scale matrix.
fn scale(s: f32) -> Self {
Self {
a: s,
b: 0.0,
c: 0.0,
d: 0.0,
e: s,
f: 0.0,
}
}
/// Compose this transformation with another (self * other).
/// This applies 'other' first, then 'self'.
fn compose(&self, other: &AffineMatrix) -> AffineMatrix {
AffineMatrix {
a: self.a * other.a + self.b * other.d,
b: self.a * other.b + self.b * other.e,
c: self.a * other.c + self.b * other.f + self.c,
d: self.d * other.a + self.e * other.d,
e: self.d * other.b + self.e * other.e,
f: self.d * other.c + self.e * other.f + self.f,
}
}
/// Transform a point using this matrix.
fn transform_point(&self, x: f32, y: f32) -> (f32, f32) {
(
self.a * x + self.b * y + self.c,
self.d * x + self.e * y + self.f,
)
}
}
/// Calculate the affine transformation matrix to align eyes at desired positions.
///
/// This implements the same algorithm as the Python example:
/// 1. Calculate target eye positions based on config
/// 2. Compute rotation angle to make eyes horizontal
/// 3. Compute scale to match desired inter-eye distance
/// 4. Combine translation, rotation, scale, and final translation into one matrix
fn calculate_eye_alignment_transform(
left_eye: Point,
right_eye: Point,
output_size: u32,
alignment_config: &crate::config::AlignmentConfig,
) -> AffineMatrix {
let output_size_f = output_size as f32;
// Calculate target eye positions from the single eye_distance setting
let left_eye_x = alignment_config.left_eye_x();
let left_eye_y = alignment_config.left_eye_y();
let left_eye_target = Point::new(output_size_f * left_eye_x, output_size_f * left_eye_y);
let right_eye_target = Point::new(
output_size_f * (1.0 - left_eye_x),
output_size_f * left_eye_y,
);
// Calculate angles
let current_angle = (right_eye.y - left_eye.y).atan2(right_eye.x - left_eye.x);
let target_angle =
(right_eye_target.y - left_eye_target.y).atan2(right_eye_target.x - left_eye_target.x);
let rotation_angle = target_angle - current_angle;
// Calculate scale
let current_eye_distance =
((right_eye.x - left_eye.x).powi(2) + (right_eye.y - left_eye.y).powi(2)).sqrt();
let target_eye_distance = ((right_eye_target.x - left_eye_target.x).powi(2)
+ (right_eye_target.y - left_eye_target.y).powi(2))
.sqrt();
let scale = target_eye_distance / current_eye_distance;
// Eye centers
let center = Point::new(
(left_eye.x + right_eye.x) / 2.0,
(left_eye.y + right_eye.y) / 2.0,
);
let target_center = Point::new(
(left_eye_target.x + right_eye_target.x) / 2.0,
(left_eye_target.y + right_eye_target.y) / 2.0,
);
// Build the transformation matrix by composing:
// 1. Translate to origin (center of eyes)
// 2. Rotate
// 3. Scale
// 4. Translate to target position
let m1 = AffineMatrix::translation(-center.x, -center.y);
let m2 = AffineMatrix::rotation(rotation_angle);
let m3 = AffineMatrix::scale(scale);
let m4 = AffineMatrix::translation(target_center.x, target_center.y);
// Compose: M = M4 * M3 * M2 * M1
// This means we apply M1 first, then M2, then M3, then M4
m4.compose(&m3.compose(&m2.compose(&m1)))
}
/// Apply an affine transformation to an image.
///
/// Uses inverse mapping with bilinear interpolation to avoid holes in the output.
fn apply_affine_transform(
image: &DynamicImage,
transform: &AffineMatrix,
output_size: u32,
) -> DynamicImage {
let rgb = image.to_rgb8();
let (width, height) = rgb.dimensions();
// We need the inverse transform to do inverse mapping
// For an affine transform, the inverse can be computed analytically
let det = transform.a * transform.e - transform.b * transform.d;
if det.abs() < 1e-10 {
// Degenerate transform, return black image
return DynamicImage::ImageRgb8(RgbImage::new(output_size, output_size));
}
let inv_det = 1.0 / det;
let inv_transform = AffineMatrix {
a: transform.e * inv_det,
b: -transform.b * inv_det,
c: (transform.b * transform.f - transform.e * transform.c) * inv_det,
d: -transform.d * inv_det,
e: transform.a * inv_det,
f: (transform.d * transform.c - transform.a * transform.f) * inv_det,
};
// Create output image using inverse mapping
let output = RgbImage::from_fn(output_size, output_size, |x, y| {
// Map output pixel to source pixel
let (src_x, src_y) = inv_transform.transform_point(x as f32, y as f32);
// Clamp coordinates to valid range (replicate edge pixels)
let clamped_x = src_x.clamp(0.0, (width - 1) as f32);
let clamped_y = src_y.clamp(0.0, (height - 1) as f32);
// Bilinear interpolation with clamped coordinates
let x0 = clamped_x.floor() as u32;
let y0 = clamped_y.floor() as u32;
let x1 = (x0 + 1).min(width - 1);
let y1 = (y0 + 1).min(height - 1);
let dx = clamped_x - x0 as f32;
let dy = clamped_y - y0 as f32;
let p00 = rgb.get_pixel(x0, y0);
let p10 = rgb.get_pixel(x1, y0);
let p01 = rgb.get_pixel(x0, y1);
let p11 = rgb.get_pixel(x1, y1);
let r = (p00[0] as f32 * (1.0 - dx) * (1.0 - dy)
+ p10[0] as f32 * dx * (1.0 - dy)
+ p01[0] as f32 * (1.0 - dx) * dy
+ p11[0] as f32 * dx * dy)
.round()
.clamp(0.0, 255.0) as u8;
let g = (p00[1] as f32 * (1.0 - dx) * (1.0 - dy)
+ p10[1] as f32 * dx * (1.0 - dy)
+ p01[1] as f32 * (1.0 - dx) * dy
+ p11[1] as f32 * dx * dy)
.round()
.clamp(0.0, 255.0) as u8;
let b = (p00[2] as f32 * (1.0 - dx) * (1.0 - dy)
+ p10[2] as f32 * dx * (1.0 - dy)
+ p01[2] as f32 * (1.0 - dx) * dy
+ p11[2] as f32 * dx * dy)
.round()
.clamp(0.0, 255.0) as u8;
Rgb([r, g, b])
});
DynamicImage::ImageRgb8(output)
}
#[cfg(test)]
mod tests {
use super::*;
use crate::immich_api::FaceData;
use image::RgbImage;
fn make_test_ctx() -> PipelineContext {
let face_data = FaceData {
bounding_box_x1: 0.0,
bounding_box_y1: 0.0,
bounding_box_x2: 100.0,
bounding_box_y2: 100.0,
image_width: 100,
image_height: 100,
};
PipelineContext::new("test".to_string(), "2024-01-01".to_string(), face_data)
}
#[tokio::test]
async fn test_no_landmarks_errors() {
let step = AlignmentStep;
let ctx = make_test_ctx();
let config = Config::default();
// Create a dummy image but no landmarks
let img = DynamicImage::ImageRgb8(RgbImage::new(100, 100));
let ctx = ctx.with_image(img);
match step.execute(ctx, &config).await {
StepOutcome::Error { error, .. } => {
assert!(error.contains("Landmarks not available"));
}
other => panic!("Expected Error without landmarks, got {:?}", other),
}
}
}