immich-automated-selfie-tim.../src/pipeline/steps/head_pose.rs
2026-02-12 21:46:30 +01:00

384 lines
12 KiB
Rust

//! Head pose estimation step.
//!
//! Uses the DMHead ONNX model to estimate head pose (yaw, pitch, roll) and
//! filter out non-front-facing faces.
use crate::config::Config;
use crate::models::DMHeadModel;
use crate::pipeline::{
computed_keys, draw_simple_text, ComputedValue, PipelineContext, ProcessingStep, StepOutcome,
};
use async_trait::async_trait;
use image::{DynamicImage, GenericImageView, Rgb, RgbImage};
/// 3D point for cube vertices
#[derive(Clone, Copy)]
struct Point3D {
x: f32,
y: f32,
z: f32,
}
impl Point3D {
fn new(x: f32, y: f32, z: f32) -> Self {
Self { x, y, z }
}
/// Rotate around Y axis (yaw)
fn rotate_y(self, angle_deg: f32) -> Self {
let rad = angle_deg.to_radians();
let cos = rad.cos();
let sin = rad.sin();
Self {
x: self.x * cos + self.z * sin,
y: self.y,
z: -self.x * sin + self.z * cos,
}
}
/// Rotate around X axis (pitch)
fn rotate_x(self, angle_deg: f32) -> Self {
let rad = angle_deg.to_radians();
let cos = rad.cos();
let sin = rad.sin();
Self {
x: self.x,
y: self.y * cos - self.z * sin,
z: self.y * sin + self.z * cos,
}
}
/// Rotate around Z axis (roll)
fn rotate_z(self, angle_deg: f32) -> Self {
let rad = angle_deg.to_radians();
let cos = rad.cos();
let sin = rad.sin();
Self {
x: self.x * cos - self.y * sin,
y: self.x * sin + self.y * cos,
z: self.z,
}
}
/// Project 3D point to 2D using perspective projection
fn project(self, cx: f32, cy: f32, focal_length: f32) -> (i32, i32) {
let scale = focal_length / (focal_length + self.z);
let x2d = cx + self.x * scale;
let y2d = cy + self.y * scale;
(x2d as i32, y2d as i32)
}
}
/// Draw 3D pose axes (RGB = XYZ) from face center, rotated by head pose.
/// This is the standard visualization for head pose estimation.
fn draw_pose_axes(
img: &mut RgbImage,
cx: f32,
cy: f32,
axis_length: f32,
yaw: f32,
pitch: f32,
roll: f32,
) {
// Define axis endpoints (origin at 0,0,0)
// X axis (red) - points right
// Y axis (green) - points down (image coordinates)
// Z axis (blue) - points out of screen (towards camera)
let axes = [
(Point3D::new(axis_length, 0.0, 0.0), Rgb([255, 0, 0])), // X - red
(Point3D::new(0.0, axis_length, 0.0), Rgb([0, 255, 0])), // Y - green
(Point3D::new(0.0, 0.0, -axis_length), Rgb([0, 0, 255])), // Z - blue (negative = towards camera)
];
// Negate roll and pitch to convert from model convention to image coordinates
// (image Y-axis points down, model assumes Y-axis points up)
let focal_length = axis_length * 2.0;
let origin = Point3D::new(0.0, 0.0, 0.0)
.rotate_y(yaw)
.rotate_x(-pitch)
.rotate_z(-roll)
.project(cx, cy, focal_length);
for (endpoint, color) in axes {
let rotated = endpoint
.rotate_y(yaw)
.rotate_x(-pitch)
.rotate_z(-roll)
.project(cx, cy, focal_length);
draw_line(img, origin.0, origin.1, rotated.0, rotated.1, color);
}
}
/// Draw an axis-aligned bounding box (no rotation applied).
fn draw_axis_aligned_rect(img: &mut RgbImage, x1: i32, y1: i32, x2: i32, y2: i32, color: Rgb<u8>) {
draw_line(img, x1, y1, x2, y1, color); // top
draw_line(img, x2, y1, x2, y2, color); // right
draw_line(img, x2, y2, x1, y2, color); // bottom
draw_line(img, x1, y2, x1, y1, color); // left
}
/// Draw a line using Bresenham's algorithm.
fn draw_line(img: &mut RgbImage, x0: i32, y0: i32, x1: i32, y1: i32, color: Rgb<u8>) {
let (width, height) = (img.width() as i32, img.height() as i32);
let dx = (x1 - x0).abs();
let dy = -(y1 - y0).abs();
let sx = if x0 < x1 { 1 } else { -1 };
let sy = if y0 < y1 { 1 } else { -1 };
let mut err = dx + dy;
let mut x = x0;
let mut y = y0;
loop {
if x >= 0 && x < width && y >= 0 && y < height {
img.put_pixel(x as u32, y as u32, color);
}
if x == x1 && y == y1 {
break;
}
let e2 = 2 * err;
if e2 >= dy {
err += dy;
x += sx;
}
if e2 <= dx {
err += dx;
y += sy;
}
}
}
/// Estimates head pose and filters non-frontal faces.
///
/// This step:
/// 1. Runs DMHead inference on the cropped face image
/// 2. Stores the HeadPose result in ctx.computed["head_pose"]
/// 3. Skips if any angle exceeds configured thresholds
pub struct HeadPoseStep;
#[async_trait]
impl ProcessingStep for HeadPoseStep {
fn id(&self) -> &'static str {
"head_pose"
}
fn name(&self) -> &'static str {
"Head Pose"
}
async fn execute(&self, mut ctx: PipelineContext, config: &Config) -> StepOutcome {
let image: &DynamicImage = match ctx.require_image("head pose estimation") {
Ok(img) => img,
Err(e) => return StepOutcome::Error { ctx, error: e },
};
// Load the DMHead model
let model = match DMHeadModel::global() {
Ok(m) => m,
Err(e) => {
// If model isn't available, skip this step with a warning
tracing::warn!(
"DMHead model not available, skipping head pose check: {}",
e
);
return StepOutcome::Continue(ctx);
}
};
// Extract a square face crop if we have the face rectangle
// Square crop prevents aspect ratio distortion when DMHead resizes to 224x224
let face_image: DynamicImage = if let Some(face_rect) = ctx
.get_computed(computed_keys::FACE_RECT)
.and_then(|v| v.as_face_rect())
{
// Use the face rectangle to extract a tighter crop
let (img_w, img_h) = image.dimensions();
let x = (face_rect.x1 as u32).min(img_w.saturating_sub(1));
let y = (face_rect.y1 as u32).min(img_h.saturating_sub(1));
let w = ((face_rect.x2 - face_rect.x1) as u32).min(img_w - x);
let h = ((face_rect.y2 - face_rect.y1) as u32).min(img_h - y);
if w > 10 && h > 10 {
// Add a small margin around the face for better model performance
let margin = (w.max(h) / 4).min(20);
let x_with_margin = x.saturating_sub(margin);
let y_with_margin = y.saturating_sub(margin);
let w_with_margin = (w + margin * 2).min(img_w - x_with_margin);
let h_with_margin = (h + margin * 2).min(img_h - y_with_margin);
// Make the crop square by using the larger dimension
let size = w_with_margin.max(h_with_margin);
// Center the square crop around the face
let center_x = x_with_margin + w_with_margin / 2;
let center_y = y_with_margin + h_with_margin / 2;
let square_x = center_x.saturating_sub(size / 2);
let square_y = center_y.saturating_sub(size / 2);
// Ensure the square crop doesn't go out of bounds
let final_x = square_x.min(img_w.saturating_sub(size));
let final_y = square_y.min(img_h.saturating_sub(size));
let final_size = size.min(img_w - final_x).min(img_h - final_y);
image.crop_imm(final_x, final_y, final_size, final_size)
} else {
// Face rect too small, use full image
image.clone()
}
} else {
// No face rect available, use full image
image.clone()
};
// Run inference on the face crop
let pose = match model.estimate(&face_image) {
Ok(p) => p,
Err(e) => {
return StepOutcome::Error {
ctx,
error: format!("Head pose estimation failed: {}", e),
};
}
};
// Store pose in computed values
ctx.set_computed(computed_keys::HEAD_POSE, ComputedValue::HeadPose(pose));
// Check against thresholds
let head_pose_config = &config.processing.head_pose;
if pose.yaw.abs() > head_pose_config.max_yaw {
return StepOutcome::Skip {
ctx,
reason: "head_turned".to_string(),
detail: Some(format!(
"Yaw {:.1}° exceeds threshold {:.1}°",
pose.yaw, head_pose_config.max_yaw
)),
};
}
if pose.pitch.abs() > head_pose_config.max_pitch {
return StepOutcome::Skip {
ctx,
reason: "head_turned".to_string(),
detail: Some(format!(
"Pitch {:.1}° exceeds threshold {:.1}°",
pose.pitch, head_pose_config.max_pitch
)),
};
}
// Note: Roll is not checked for pass/fail - only used for visualization
// Roll (head tilt) is less important for timelapse alignment since
// the alignment step can handle rotated faces
StepOutcome::Continue(ctx)
}
fn debug_visualize(&self, ctx: &PipelineContext, _config: &Config) -> Option<DynamicImage> {
// Get head pose from computed values
let pose = ctx
.get_computed(computed_keys::HEAD_POSE)
.and_then(|v| v.as_head_pose())?;
// Get the current image to draw on
let image = ctx.image.as_ref()?;
let rgb = image.to_rgb8();
let (width, height) = (rgb.width(), rgb.height());
// Create a copy for visualization
let mut debug_img = rgb.clone();
// Get face rectangle if available
let face_rect = ctx
.get_computed(computed_keys::FACE_RECT)
.and_then(|v| v.as_face_rect());
// Draw axis-aligned face bounding box and 3D pose axes
if let Some(rect) = face_rect {
let x1 = rect.x1 as i32;
let y1 = rect.y1 as i32;
let x2 = rect.x2 as i32;
let y2 = rect.y2 as i32;
// Draw axis-aligned bounding box in cyan
draw_axis_aligned_rect(&mut debug_img, x1, y1, x2, y2, Rgb([0, 255, 255]));
// Draw 3D pose axes from face center
let cx = (x1 + x2) as f32 / 2.0;
let cy = (y1 + y2) as f32 / 2.0;
let axis_length = ((x2 - x1).max(y2 - y1) as f32) * 0.6;
draw_pose_axes(
&mut debug_img,
cx,
cy,
axis_length,
pose.yaw,
pose.pitch,
pose.roll,
);
}
// Draw text background bar at bottom for pose values
let bar_height = 20u32;
for y in height.saturating_sub(bar_height)..height {
for x in 0..width {
debug_img.put_pixel(x, y, Rgb([0, 0, 0]));
}
}
// Draw simple text representation of values
let text = format!(
"Y:{:+.0} P:{:+.0} R:{:+.0}",
pose.yaw, pose.pitch, pose.roll
);
draw_simple_text(
&mut debug_img,
5,
height - bar_height + 4,
&text,
Rgb([255, 255, 255]),
);
Some(DynamicImage::ImageRgb8(debug_img))
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::immich_api::FaceData;
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_image_error() {
let step = HeadPoseStep;
let ctx = make_test_ctx();
let mut config = Config::default();
config.processing.head_pose.enabled = true;
match step.execute(ctx, &config).await {
StepOutcome::Error { error, .. } => {
assert!(error.contains("No image"));
}
other => panic!("Expected Error, got {:?}", other),
}
}
}