oxedyne/fe2o3/fe2o3_infer/src/face/mod.rs
5.5 KiB, 1 run
created by r1870400018:19753, which is this file's identity for as long as the history lasts, whatever it is later renamed to
download · who wrote it · its history
| 1 | //! Face detection and face embedding: the two things this crate exists to do. |
| 2 | //! |
| 3 | //! # Order of operations |
| 4 | //! |
| 5 | //! 1. Fit the photograph into the detector's canvas with [`letterbox`], which |
| 6 | //! answers the scale needed to put the results back in the original frame. |
| 7 | //! 2. Run [`Detector::detect`], which gives a box, a score and five landmarks |
| 8 | //! per face. |
| 9 | //! 3. Run [`Embedder::embed`] on the original photograph and one detection's |
| 10 | //! landmarks, which warps the face onto a fixed template and answers a |
| 11 | //! hundred and twenty-eight dimensional unit vector. |
| 12 | //! 4. Compare two of those with [`cosine`]. |
| 13 | //! |
| 14 | //! # Channel order |
| 15 | //! |
| 16 | //! The two networks disagree, and neither says so. The detector was exported |
| 17 | //! against blue-green-red input and the embedder against red-green-blue. Both |
| 18 | //! entry points here take ordinary red-green-blue pixels and put the channels |
| 19 | //! in the order each network was trained on, so a caller never has to know. |
| 20 | |
| 21 | pub mod align; |
| 22 | pub mod detect; |
| 23 | pub mod embed; |
| 24 | |
| 25 | pub use align::{ |
| 26 | align_crop, |
| 27 | similarity, |
| 28 | Affine, |
| 29 | CROP, |
| 30 | TEMPLATE, |
| 31 | }; |
| 32 | pub use detect::{ |
| 33 | Detection, |
| 34 | Detector, |
| 35 | DetectorOptions, |
| 36 | }; |
| 37 | pub use embed::{ |
| 38 | cosine, |
| 39 | Embedder, |
| 40 | Embedding, |
| 41 | }; |
| 42 | |
| 43 | use oxedyne_fe2o3_core::prelude::*; |
| 44 | |
| 45 | /// A borrowed, interleaved, eight-bit image. |
| 46 | #[derive(Clone, Copy, Debug)] |
| 47 | pub struct Image<'a> { |
| 48 | /// Pixels, row-major, `channels` values per pixel. |
| 49 | pub pixels: &'a [u8], |
| 50 | /// Width in pixels. |
| 51 | pub width: usize, |
| 52 | /// Height in pixels. |
| 53 | pub height: usize, |
| 54 | /// Values per pixel, three for red-green-blue. |
| 55 | pub channels: usize, |
| 56 | } |
| 57 | |
| 58 | impl<'a> Image<'a> { |
| 59 | /// Wraps a buffer, checking that it holds what the extents claim. |
| 60 | pub fn new(pixels: &'a [u8], width: usize, height: usize, channels: usize) |
| 61 | -> Outcome<Self> |
| 62 | { |
| 63 | let want = width * height * channels; |
| 64 | if pixels.len() != want { |
| 65 | return Err(err!( |
| 66 | "An image of {} by {} with {} channels wants {} bytes, but {} were given.", |
| 67 | width, height, channels, want, pixels.len(); |
| 68 | Invalid, Input, Mismatch)); |
| 69 | } |
| 70 | Ok(Self { pixels, width, height, channels }) |
| 71 | } |
| 72 | |
| 73 | /// Reads one channel of one pixel, answering zero outside the frame, which |
| 74 | /// is the constant border a warp needs. |
| 75 | #[inline] |
| 76 | pub fn sample(&self, x: f64, y: f64, c: usize) -> f64 { |
| 77 | if x < 0.0 || y < 0.0 || c >= self.channels { |
| 78 | return 0.0; |
| 79 | } |
| 80 | let (xi, yi) = (x as usize, y as usize); |
| 81 | if xi >= self.width || yi >= self.height { |
| 82 | return 0.0; |
| 83 | } |
| 84 | self.pixels[(yi * self.width + xi) * self.channels + c] as f64 |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | /// What a letterbox did, so that a result can be put back in the original frame. |
| 89 | #[derive(Clone, Copy, Debug, PartialEq)] |
| 90 | pub struct Letterbox { |
| 91 | /// Multiplier applied to the original, at most one. |
| 92 | pub scale: f64, |
| 93 | /// Canvas width. |
| 94 | pub width: usize, |
| 95 | /// Canvas height. |
| 96 | pub height: usize, |
| 97 | } |
| 98 | |
| 99 | impl Letterbox { |
| 100 | /// Maps a point on the canvas back to the original frame. |
| 101 | pub fn back(&self, x: f32, y: f32) -> (f32, f32) { |
| 102 | ((x as f64 / self.scale) as f32, (y as f64 / self.scale) as f32) |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | /// Fits an image into a canvas of the given size, keeping the aspect ratio and |
| 107 | /// leaving the unused right and lower margin black. |
| 108 | /// |
| 109 | /// Downscaling averages over the source footprint rather than taking one sample |
| 110 | /// from it, because a face sixty pixels across in a four thousand pixel |
| 111 | /// photograph is ten pixels across in a six hundred and forty pixel canvas, and |
| 112 | /// a point sample of it is noise. |
| 113 | pub fn letterbox(img: &Image<'_>, width: usize, height: usize) |
| 114 | -> Outcome<(Vec<u8>, Letterbox)> |
| 115 | { |
| 116 | if width == 0 || height == 0 || img.width == 0 || img.height == 0 { |
| 117 | return Err(err!( |
| 118 | "A letterbox of {} by {} from {} by {} has no area.", |
| 119 | width, height, img.width, img.height; |
| 120 | Invalid, Input, Range)); |
| 121 | } |
| 122 | let ch = img.channels; |
| 123 | let scale = (width as f64 / img.width as f64).min(height as f64 / img.height as f64); |
| 124 | let dw = ((img.width as f64 * scale).round() as usize).clamp(1, width); |
| 125 | let dh = ((img.height as f64 * scale).round() as usize).clamp(1, height); |
| 126 | let mut out = vec![0u8; width * height * ch]; |
| 127 | |
| 128 | if scale <= 1.0 { |
| 129 | // Area average: each destination pixel is the mean of the source |
| 130 | // rectangle that maps onto it. |
| 131 | for y in 0..dh { |
| 132 | let y0 = (y as f64 * img.height as f64 / dh as f64).floor() as usize; |
| 133 | let y1 = (((y + 1) as f64 * img.height as f64 / dh as f64).ceil() as usize) |
| 134 | .clamp(y0 + 1, img.height); |
| 135 | for x in 0..dw { |
| 136 | let x0 = (x as f64 * img.width as f64 / dw as f64).floor() as usize; |
| 137 | let x1 = (((x + 1) as f64 * img.width as f64 / dw as f64).ceil() as usize) |
| 138 | .clamp(x0 + 1, img.width); |
| 139 | let n = ((y1 - y0) * (x1 - x0)) as f64; |
| 140 | for c in 0..ch { |
| 141 | let mut s = 0.0f64; |
| 142 | for sy in y0..y1 { |
| 143 | for sx in x0..x1 { |
| 144 | s += img.pixels[(sy * img.width + sx) * ch + c] as f64; |
| 145 | } |
| 146 | } |
| 147 | out[(y * width + x) * ch + c] = (s / n).round().clamp(0.0, 255.0) as u8; |
| 148 | } |
| 149 | } |
| 150 | } |
| 151 | } else { |
| 152 | // Bilinear, which is what an enlargement wants. |
| 153 | for y in 0..dh { |
| 154 | let sy = (y as f64 + 0.5) / scale - 0.5; |
| 155 | let yb = sy.floor(); |
| 156 | let fy = sy - yb; |
| 157 | for x in 0..dw { |
| 158 | let sx = (x as f64 + 0.5) / scale - 0.5; |
| 159 | let xb = sx.floor(); |
| 160 | let fx = sx - xb; |
| 161 | for c in 0..ch { |
| 162 | let p00 = img.sample(xb.max(0.0), yb.max(0.0), c); |
| 163 | let p10 = img.sample((xb + 1.0).max(0.0), yb.max(0.0), c); |
| 164 | let p01 = img.sample(xb.max(0.0), (yb + 1.0).max(0.0), c); |
| 165 | let p11 = img.sample((xb + 1.0).max(0.0), (yb + 1.0).max(0.0), c); |
| 166 | let top = p00 + (p10 - p00) * fx; |
| 167 | let bot = p01 + (p11 - p01) * fx; |
| 168 | let v = top + (bot - top) * fy; |
| 169 | out[(y * width + x) * ch + c] = v.round().clamp(0.0, 255.0) as u8; |
| 170 | } |
| 171 | } |
| 172 | } |
| 173 | } |
| 174 | Ok((out, Letterbox { scale, width, height })) |
| 175 | } |