Do 2nd pass of blazeface on close ups

for better accuracy
Transform utils

src/components/PeopleList.tsx

@@ -14,14 +14,14 @@ const FaceChipContainer = styled.div`
     flex-wrap: wrap;
     justify-content: center;
     align-items: center;
-    margin-top: 10px;
-    margin-bottom: 10px;
+    margin-top: 5px;
+    margin-bottom: 5px;
 `;
 
 const FaceChip = styled.div`
     width: 112px;
     height: 112px;
-    margin-right: 10px;
+    margin: 5px;
     border-radius: 50%;
     overflow: hidden;
     position: relative;

src/services/machineLearning/tfjsFaceDetectionService.ts

@@ -10,6 +10,7 @@ import {
     BLAZEFACE_INPUT_SIZE,
     BLAZEFACE_IOU_THRESHOLD,
     BLAZEFACE_MAX_FACES,
+    BLAZEFACE_PASS1_SCORE_THRESHOLD,
     BLAZEFACE_SCORE_THRESHOLD,
     DetectedFace,
     FaceDetectionMethod,
@@ -17,7 +18,13 @@ import {
     Versioned,
 } from 'types/machineLearning';
 import { Box, Point } from '../../../thirdparty/face-api/classes';
-import { resizeToSquare } from 'utils/image';
+import { addPadding, crop, resizeToSquare } from 'utils/image';
+import {
+    computeTransformToBox,
+    transformBox,
+    transformPoints,
+} from 'utils/machineLearning/transform';
+import { enlargeBox, newBox, normFaceBox } from 'utils/machineLearning';
 
 class TFJSFaceDetectionService implements FaceDetectionService {
     private blazeFaceModel: Promise<BlazeFaceModel>;
@@ -38,7 +45,7 @@ class TFJSFaceDetectionService implements FaceDetectionService {
     private async init() {
         this.blazeFaceModel = blazeFaceLoad({
             maxFaces: BLAZEFACE_MAX_FACES,
-            scoreThreshold: BLAZEFACE_SCORE_THRESHOLD,
+            scoreThreshold: BLAZEFACE_PASS1_SCORE_THRESHOLD,
             iouThreshold: BLAZEFACE_IOU_THRESHOLD,
             modelUrl: '/models/blazeface/back/model.json',
             inputHeight: BLAZEFACE_INPUT_SIZE,
@@ -149,42 +156,79 @@ class TFJSFaceDetectionService implements FaceDetectionService {
         return this.blazeFaceModel;
     }
 
-    public async detectFaces(
+    private async estimateFaces(
         imageBitmap: ImageBitmap
     ): Promise<Array<DetectedFace>> {
         const resized = resizeToSquare(imageBitmap, BLAZEFACE_INPUT_SIZE);
-        const widthRatio = imageBitmap.width / resized.width;
-        const heightRatio = imageBitmap.height / resized.height;
         const tfImage = tf.browser.fromPixels(resized.image);
         const blazeFaceModel = await this.getBlazefaceModel();
         const faces = await blazeFaceModel.estimateFaces(tfImage);
         tf.dispose(tfImage);
 
-        const detectedFaces: Array<DetectedFace> = faces?.map(
-            (normalizedFace) => {
-                const landmarks = normalizedFace.landmarks as number[][];
+        const inBox = newBox(0, 0, resized.width, resized.height);
+        const toBox = newBox(0, 0, imageBitmap.width, imageBitmap.height);
+        const transform = computeTransformToBox(inBox, toBox);
+        // console.log("1st pass: ", { transform });
+
+        const detectedFaces: Array<DetectedFace> = faces?.map((f) => {
+            const box = transformBox(normFaceBox(f), transform);
+            const normLandmarks = (f.landmarks as number[][])?.map(
+                (l) => new Point(l[0], l[1])
+            );
+            const landmarks = transformPoints(normLandmarks, transform);
             return {
-                    box: new Box({
-                        left: normalizedFace.topLeft[0] * widthRatio,
-                        top: normalizedFace.topLeft[1] * heightRatio,
-                        right: normalizedFace.bottomRight[0] * widthRatio,
-                        bottom: normalizedFace.bottomRight[1] * heightRatio,
-                    }),
-                    landmarks:
-                        landmarks &&
-                        landmarks.map(
-                            (l) =>
-                                new Point(l[0] * widthRatio, l[1] * heightRatio)
-                        ),
-                    probability: normalizedFace.probability as number,
+                box,
+                landmarks,
+                probability: f.probability as number,
                 // detectionMethod: this.method,
             } as DetectedFace;
-            }
-        );
+        });
 
         return detectedFaces;
     }
 
+    public async detectFaces(
+        imageBitmap: ImageBitmap
+    ): Promise<Array<DetectedFace>> {
+        const pass1Detections = await this.estimateFaces(imageBitmap);
+
+        // run 2nd pass for accuracy
+        const detections = [];
+        for (const face of pass1Detections) {
+            const imageBox = enlargeBox(face.box, 2);
+            const faceImage = crop(
+                imageBitmap,
+                imageBox,
+                BLAZEFACE_INPUT_SIZE / 2
+            );
+            const paddedImage = addPadding(faceImage, 0.5);
+            const paddedBox = enlargeBox(imageBox, 2);
+            const pass2Detections = await this.estimateFaces(paddedImage);
+            // TODO: select based on nearest under certain threshold
+            // this will matter based on our IOU threshold and faces near each other
+            const selected = pass2Detections[0];
+            // console.log("pass2: ", face.probability, selected.probability);
+
+            // TODO: we might miss 1st pass face actually having score within threshold
+            if (selected && selected.probability >= BLAZEFACE_SCORE_THRESHOLD) {
+                const inBox = newBox(0, 0, faceImage.width, faceImage.height);
+                imageBox.x = paddedBox.x;
+                imageBox.y = paddedBox.y;
+                const transform = computeTransformToBox(inBox, imageBox);
+
+                selected.box = transformBox(selected.box, transform);
+                selected.landmarks = transformPoints(
+                    selected.landmarks,
+                    transform
+                );
+                // console.log("pass2: ", { imageBox, paddedBox, transform, selected });
+                detections.push(selected);
+            }
+        }
+
+        return detections;
+    }
+
     public async dispose() {
         const blazeFaceModel = await this.getBlazefaceModel();
         blazeFaceModel?.dispose();

src/types/machineLearning/index.ts

@@ -273,7 +273,8 @@ export declare type MLIndex = 'files' | 'people';
 export const BLAZEFACE_MAX_FACES = 20;
 export const BLAZEFACE_INPUT_SIZE = 256;
 export const BLAZEFACE_IOU_THRESHOLD = 0.3;
-export const BLAZEFACE_SCORE_THRESHOLD = 0.65;
+export const BLAZEFACE_SCORE_THRESHOLD = 0.7;
+export const BLAZEFACE_PASS1_SCORE_THRESHOLD = 0.4;
 export const BLAZEFACE_FACE_SIZE = 112;
 
 export interface FaceDetectionService {

src/utils/image/index.ts

@@ -39,6 +39,15 @@ export function transform(
     return offscreen.transferToImageBitmap();
 }
 
+export function crop(imageBitmap: ImageBitmap, cropBox: Box, size: number) {
+    const dimensions: Dimensions = {
+        width: size,
+        height: size,
+    };
+
+    return cropWithRotation(imageBitmap, cropBox, 0, dimensions, dimensions);
+}
+
 export function cropWithRotation(
     imageBitmap: ImageBitmap,
     cropBox: Box,
@@ -105,6 +114,24 @@ export function cropWithRotation(
     return offscreen.transferToImageBitmap();
 }
 
+export function addPadding(image: ImageBitmap, padding: number) {
+    const scale = 1 + padding * 2;
+    const width = scale * image.width;
+    const height = scale * image.height;
+    const offscreen = new OffscreenCanvas(width, height);
+    const ctx = offscreen.getContext('2d');
+    ctx.imageSmoothingEnabled = false;
+    ctx.drawImage(
+        image,
+        width / 2 - image.width / 2,
+        height / 2 - image.height / 2,
+        image.width,
+        image.height
+    );
+
+    return offscreen.transferToImageBitmap();
+}
+
 export async function imageBitmapToBlob(
     imageBitmap: ImageBitmap,
     options?: BlobOptions

src/utils/machineLearning/faceCrop.ts

@@ -40,25 +40,25 @@ export async function getStoredFaceCrop(
     };
 }
 
-export async function ibExtractFaceImageFromCrop(
-    alignedFace: AlignedFace,
+export function extractFaceImageFromCrop(
+    faceCrop: FaceCrop,
+    box: Box,
+    rotation: number,
     faceSize: number
-): Promise<ImageBitmap> {
-    const image = alignedFace.faceCrop?.image;
-    let imageBox = alignedFace.faceCrop?.imageBox;
-    if (!image || !imageBox) {
+): ImageBitmap {
+    const faceCropImage = faceCrop?.image;
+    let imageBox = faceCrop?.imageBox;
+    if (!faceCropImage || !imageBox) {
         throw Error('Face crop not present');
     }
 
-    const box = getAlignedFaceBox(alignedFace);
-    const faceCropImage = await createImageBitmap(alignedFace.faceCrop.image);
-
     // TODO: Have better serialization to avoid creating new object manually when calling class methods
     imageBox = new Box(imageBox);
     const scale = faceCropImage.width / imageBox.width;
-    const scaledImageBox = imageBox.rescale(scale).round();
-    const scaledBox = box.rescale(scale).round();
-    const shiftedBox = scaledBox.shift(-scaledImageBox.x, -scaledImageBox.y);
+    const transformedBox = box
+        .shift(-imageBox.x, -imageBox.y)
+        .rescale(scale)
+        .round();
     // console.log({ box, imageBox, faceCropImage, scale, scaledBox, scaledImageBox, shiftedBox });
 
     const faceSizeDimentions: Dimensions = {
@@ -67,8 +67,8 @@ export async function ibExtractFaceImageFromCrop(
     };
     const faceImage = cropWithRotation(
         faceCropImage,
-        shiftedBox,
-        alignedFace.rotation,
+        transformedBox,
+        rotation,
         faceSizeDimentions,
         faceSizeDimentions
     );
@@ -76,6 +76,21 @@ export async function ibExtractFaceImageFromCrop(
     return faceImage;
 }
 
+export async function ibExtractFaceImageFromCrop(
+    alignedFace: AlignedFace,
+    faceSize: number
+): Promise<ImageBitmap> {
+    const box = getAlignedFaceBox(alignedFace);
+    const faceCropImage = await createImageBitmap(alignedFace.faceCrop.image);
+
+    return extractFaceImageFromCrop(
+        { image: faceCropImage, imageBox: alignedFace.faceCrop?.imageBox },
+        box,
+        alignedFace.rotation,
+        faceSize
+    );
+}
+
 export async function ibExtractFaceImagesFromCrops(
     faces: AlignedFace[],
     faceSize: number

src/utils/machineLearning/index.ts

@@ -16,6 +16,7 @@ import { mlFilesStore, mlPeopleStore } from 'utils/storage/mlStorage';
 import { convertForPreview, needsConversionForPreview } from 'utils/file';
 import { cached } from 'utils/storage/cache';
 import { imageBitmapToBlob } from 'utils/image';
+import { NormalizedFace } from '@tensorflow-models/blazeface';
 
 export function f32Average(descriptors: Float32Array[]) {
     if (descriptors.length < 1) {
@@ -123,6 +124,28 @@ export function extractFaces(
     });
 }
 
+export function newBox(x: number, y: number, width: number, height: number) {
+    return new Box({ x, y, width, height });
+}
+
+export function newBoxFromPoints(
+    left: number,
+    top: number,
+    right: number,
+    bottom: number
+) {
+    return new Box({ left, top, right, bottom });
+}
+
+export function normFaceBox(face: NormalizedFace) {
+    return newBoxFromPoints(
+        face.topLeft[0],
+        face.topLeft[1],
+        face.bottomRight[0],
+        face.bottomRight[1]
+    );
+}
+
 export function getBoxCenterPt(topLeft: Point, bottomRight: Point): Point {
     return topLeft.add(bottomRight.sub(topLeft).div(new Point(2, 2)));
 }
@@ -204,6 +227,15 @@ async function getOriginalImageFile(file: File, token: string) {
     return new Response(fileStream).blob();
 }
 
+async function getOriginalFile(file: File, token: string) {
+    let fileBlob = await getOriginalImageFile(file, token);
+    if (needsConversionForPreview(file)) {
+        fileBlob = await convertForPreview(file, fileBlob);
+    }
+
+    return fileBlob;
+}
+
 export async function getOriginalImageBitmap(
     file: File,
     token: string,
@@ -213,17 +245,13 @@ export async function getOriginalImageBitmap(
 
     if (useCache) {
         fileBlob = await cached('files', '/' + file.id.toString(), () => {
-            return getOriginalImageFile(file, token);
+            return getOriginalFile(file, token);
         });
     } else {
-        fileBlob = await getOriginalImageFile(file, token);
+        fileBlob = await getOriginalFile(file, token);
     }
     console.log('[MLService] Got file: ', file.id.toString());
 
-    if (needsConversionForPreview(file)) {
-        fileBlob = await convertForPreview(file, fileBlob);
-    }
-
     return getImageBitmap(fileBlob);
 }
 

src/utils/machineLearning/transform.ts

@@ -0,0 +1,64 @@
+import { Box, Point } from '../../../thirdparty/face-api/classes';
+import { Matrix } from 'ml-matrix';
+import { newBoxFromPoints } from '.';
+
+export function translation(x: number, y: number) {
+    return new Matrix([
+        [1, 0, x],
+        [0, 1, y],
+        [0, 0, 1],
+    ]);
+}
+
+export function scale(sx: number, sy: number) {
+    return new Matrix([
+        [sx, 0, 0],
+        [0, sy, 0],
+        [0, 0, 1],
+    ]);
+}
+
+export function rotation(angle: number) {
+    const cosa = Math.cos(angle);
+    const sina = Math.sin(angle);
+    return new Matrix([
+        [cosa, -sina, 0],
+        [sina, cosa, 0],
+        [0, 0, 1],
+    ]);
+}
+
+export function computeTransformToBox(inBox: Box, toBox: Box): Matrix {
+    return translation(toBox.x, toBox.y).mmul(
+        scale(toBox.width / inBox.width, toBox.height / inBox.height)
+    );
+}
+
+export function pointToArray(point: Point) {
+    return [point.x, point.y];
+}
+
+export function transformPointVec(point: number[], transform: Matrix) {
+    point[2] = 1;
+    const mat = new Matrix([point]).transpose();
+    const mulmat = new Matrix(transform).mmul(mat).to1DArray();
+    // console.log({point, mat, mulmat});
+
+    return mulmat;
+}
+
+export function transformPoint(point: Point, transform: Matrix) {
+    const pointVec = transformPointVec(pointToArray(point), transform);
+    return new Point(pointVec[0], pointVec[1]);
+}
+
+export function transformPoints(points: Point[], transform: Matrix) {
+    return points.map((p) => transformPoint(p, transform));
+}
+
+export function transformBox(box: Box, transform: Matrix) {
+    const topLeft = transformPoint(box.topLeft, transform);
+    const bottomRight = transformPoint(box.bottomRight, transform);
+
+    return newBoxFromPoints(topLeft.x, topLeft.y, bottomRight.x, bottomRight.y);
+}