redesign body and hand caching and interpolation

demo/index.js

@@ -277,12 +277,10 @@ async function drawResults(input) {
   }
 
   // draw all results using interpolated results
-  if (ui.interpolated) {
+  let interpolated;
-    const interpolated = human.next(result);
+  if (ui.interpolated) interpolated = human.next(result);
-    human.draw.all(canvas, interpolated, drawOptions);
+  else interpolated = result;
-  } else {
+  human.draw.all(canvas, interpolated, drawOptions);
-    human.draw.all(canvas, result, drawOptions);
-  }
 
   // show tree with results
   if (ui.results) {
@@ -315,7 +313,7 @@ async function drawResults(input) {
   document.getElementById('log').innerHTML = `
     video: ${ui.camera.name} | facing: ${ui.camera.facing} | screen: ${window.innerWidth} x ${window.innerHeight} camera: ${ui.camera.width} x ${ui.camera.height} ${processing}<br>
     backend: ${backend}<br>
-    performance: ${str(lastDetectedResult.performance)}ms ${fps}<br>
+    performance: ${str(interpolated.performance)}ms ${fps}<br>
     ${warning}<br>
   `;
   ui.framesDraw++;

src/body/blazeposecoords.ts

@@ -1,6 +1,6 @@
 /* eslint-disable no-multi-spaces */
 
-export const kpt = [
+export const kpt: Array<string> = [
   'nose',            //  0
   'leftEyeInside',   //  1
   'leftEye',         //  2
@@ -42,7 +42,7 @@ export const kpt = [
   'rightHand',       // 38
 ];
 
-export const connected = {
+export const connected: Record<string, string[]> = {
   leftLeg: ['leftHip', 'leftKnee', 'leftAnkle', 'leftHeel', 'leftFoot'],
   rightLeg: ['rightHip', 'rightKnee', 'rightAnkle', 'rightHeel', 'rightFoot'],
   torso: ['leftShoulder', 'rightShoulder', 'rightHip', 'leftHip', 'leftShoulder'],

src/body/efficientposecoords.ts

@@ -1,4 +1,4 @@
-export const kpt = [
+export const kpt: Array<string> = [
   'head',
   'neck',
   'rightShoulder',
@@ -17,7 +17,7 @@ export const kpt = [
   'leftAnkle',
 ];
 
-export const connected = {
+export const connected: Record<string, string[]> = {
   leftLeg: ['leftHip', 'leftKnee', 'leftAnkle'],
   rightLeg: ['rightHip', 'rightKnee', 'rightAnkle'],
   torso: ['leftShoulder', 'rightShoulder', 'rightHip', 'leftHip', 'leftShoulder'],

src/body/movenet.ts

@@ -5,7 +5,7 @@
  */
 
 import { log, join } from '../util/util';
-import { scale } from '../util/box';
+import * as box from '../util/box';
 import * as tf from '../../dist/tfjs.esm.js';
 import * as coords from './movenetcoords';
 import type { BodyKeypoint, BodyResult, Box, Point } from '../result';
@@ -16,7 +16,15 @@ import { env } from '../util/env';
 
 let model: GraphModel | null;
 let inputSize = 0;
-const cachedBoxes: Array<Box> = [];
+const boxExpandFact = 1.5; // increase to 150%
+
+const cache: {
+  boxes: Array<Box>,
+  bodies: Array<BodyResult>;
+} = {
+  boxes: [],
+  bodies: [],
+};
 
 let skipped = Number.MAX_SAFE_INTEGER;
 const keypoints: Array<BodyKeypoint> = [];
@@ -34,26 +42,6 @@ export async function load(config: Config): Promise<GraphModel> {
   return model;
 }
 
-function createBox(points): [Box, Box] {
-  const x = points.map((a) => a.position[0]);
-  const y = points.map((a) => a.position[1]);
-  const box: Box = [
-    Math.min(...x),
-    Math.min(...y),
-    Math.max(...x) - Math.min(...x),
-    Math.max(...y) - Math.min(...y),
-  ];
-  const xRaw = points.map((a) => a.positionRaw[0]);
-  const yRaw = points.map((a) => a.positionRaw[1]);
-  const boxRaw: Box = [
-    Math.min(...xRaw),
-    Math.min(...yRaw),
-    Math.max(...xRaw) - Math.min(...xRaw),
-    Math.max(...yRaw) - Math.min(...yRaw),
-  ];
-  return [box, boxRaw];
-}
-
 async function parseSinglePose(res, config, image, inputBox) {
   const kpt = res[0][0];
   keypoints.length = 0;
@@ -78,7 +66,7 @@ async function parseSinglePose(res, config, image, inputBox) {
   }
   score = keypoints.reduce((prev, curr) => (curr.score > prev ? curr.score : prev), 0);
   const bodies: Array<BodyResult> = [];
-  const [box, boxRaw] = createBox(keypoints);
+  const newBox = box.calc(keypoints.map((pt) => pt.position), [image.shape[2], image.shape[1]]);
   const annotations: Record<string, Point[][]> = {};
   for (const [name, indexes] of Object.entries(coords.connected)) {
     const pt: Array<Point[]> = [];
@@ -89,7 +77,7 @@ async function parseSinglePose(res, config, image, inputBox) {
     }
     annotations[name] = pt;
   }
-  bodies.push({ id: 0, score, box, boxRaw, keypoints, annotations });
+  bodies.push({ id: 0, score, box: newBox.box, boxRaw: newBox.boxRaw, keypoints, annotations });
   return bodies;
 }
 
@@ -111,14 +99,11 @@ async function parseMultiPose(res, config, image, inputBox) {
             part: coords.kpt[i],
             score: Math.round(100 * score) / 100,
             positionRaw,
-            position: [
+            position: [Math.round((image.shape[2] || 0) * positionRaw[0]), Math.round((image.shape[1] || 0) * positionRaw[1])],
-              Math.round((image.shape[2] || 0) * positionRaw[0]),
-              Math.round((image.shape[1] || 0) * positionRaw[1]),
-            ],
           });
         }
       }
-      const [box, boxRaw] = createBox(keypoints);
+      const newBox = box.calc(keypoints.map((pt) => pt.position), [image.shape[2], image.shape[1]]);
       // movenet-multipose has built-in box details
       // const boxRaw: Box = [kpt[51 + 1], kpt[51 + 0], kpt[51 + 3] - kpt[51 + 1], kpt[51 + 2] - kpt[51 + 0]];
       // const box: Box = [Math.trunc(boxRaw[0] * (image.shape[2] || 0)), Math.trunc(boxRaw[1] * (image.shape[1] || 0)), Math.trunc(boxRaw[2] * (image.shape[2] || 0)), Math.trunc(boxRaw[3] * (image.shape[1] || 0))];
@@ -132,7 +117,7 @@ async function parseMultiPose(res, config, image, inputBox) {
         }
         annotations[name] = pt;
       }
-      bodies.push({ id, score: totalScore, boxRaw, box, keypoints: [...keypoints], annotations });
+      bodies.push({ id, score: totalScore, box: newBox.box, boxRaw: newBox.boxRaw, keypoints: [...keypoints], annotations });
     }
   }
   bodies.sort((a, b) => b.score - a.score);
@@ -141,46 +126,44 @@ async function parseMultiPose(res, config, image, inputBox) {
 }
 
 export async function predict(input: Tensor, config: Config): Promise<BodyResult[]> {
-  if (!model || !model?.inputs[0].shape) return [];
+  if (!model || !model?.inputs[0].shape) return []; // something is wrong with the model
+  if (!config.skipFrame) cache.boxes.length = 0; // allowed to use cache or not
+  skipped++; // increment skip frames
+  if (config.skipFrame && (skipped <= (config.body.skipFrames || 0))) {
+    return cache.bodies; // return cached results without running anything
+  }
   return new Promise(async (resolve) => {
     const t: Record<string, Tensor> = {};
-
+    skipped = 0;
-    let bodies: Array<BodyResult> = [];
+    cache.bodies = []; // reset bodies result
-
+    if (cache.boxes.length >= (config.body.maxDetected || 0)) { // if we have enough cached boxes run detection using cache
-    if (!config.skipFrame) cachedBoxes.length = 0; // allowed to use cache or not
+      for (let i = 0; i < cache.boxes.length; i++) { // run detection based on cached boxes
-    skipped++;
+        t.crop = tf.image.cropAndResize(input, [cache.boxes[i]], [0], [inputSize, inputSize], 'bilinear');
-
+        t.cast = tf.cast(t.crop, 'int32');
-    for (let i = 0; i < cachedBoxes.length; i++) { // run detection based on cached boxes
+        t.res = await model?.predict(t.cast) as Tensor;
-      t.crop = tf.image.cropAndResize(input, [cachedBoxes[i]], [0], [inputSize, inputSize], 'bilinear');
+        const res = await t.res.array();
-      t.cast = tf.cast(t.crop, 'int32');
+        const newBodies = (t.res.shape[2] === 17) ? await parseSinglePose(res, config, input, cache.boxes[i]) : await parseMultiPose(res, config, input, cache.boxes[i]);
-      t.res = await model?.predict(t.cast) as Tensor;
+        cache.bodies = cache.bodies.concat(newBodies);
-      const res = await t.res.array();
+        Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));
-      const newBodies = (t.res.shape[2] === 17) ? await parseSinglePose(res, config, input, cachedBoxes[i]) : await parseMultiPose(res, config, input, cachedBoxes[i]);
+      }
-      bodies = bodies.concat(newBodies);
-      Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));
     }
-
+    if (cache.bodies.length !== config.body.maxDetected) { // did not find enough bodies based on cached boxes so run detection on full frame
-    if ((bodies.length !== config.body.maxDetected) && (skipped > (config.body.skipFrames || 0))) { // run detection on full frame
       t.resized = tf.image.resizeBilinear(input, [inputSize, inputSize], false);
       t.cast = tf.cast(t.resized, 'int32');
       t.res = await model?.predict(t.cast) as Tensor;
       const res = await t.res.array();
-      bodies = (t.res.shape[2] === 17) ? await parseSinglePose(res, config, input, [0, 0, 1, 1]) : await parseMultiPose(res, config, input, [0, 0, 1, 1]);
+      cache.bodies = (t.res.shape[2] === 17) ? await parseSinglePose(res, config, input, [0, 0, 1, 1]) : await parseMultiPose(res, config, input, [0, 0, 1, 1]);
+      // cache.bodies = cache.bodies.map((body) => ({ ...body, box: box.scale(body.box, 0.5) }));
       Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));
-      cachedBoxes.length = 0; // reset cache
-      skipped = 0;
     }
-
+    cache.boxes.length = 0; // reset cache
-    if (config.skipFrame) { // create box cache based on last detections
+    for (let i = 0; i < cache.bodies.length; i++) {
-      cachedBoxes.length = 0;
+      if (cache.bodies[i].keypoints.length > (coords.kpt.length / 2)) { // only update cache if we detected at least half keypoints
-      for (let i = 0; i < bodies.length; i++) {
+        const scaledBox = box.scale(cache.bodies[i].boxRaw, boxExpandFact);
-        if (bodies[i].keypoints.length > 10) { // only update cache if we detected sufficient number of keypoints
+        const cropBox = box.crop(scaledBox);
-          const kpts = bodies[i].keypoints.map((kpt) => kpt.position);
+        cache.boxes.push(cropBox);
-          const newBox = scale(kpts, 1.5, [input.shape[2], input.shape[1]]);
-          cachedBoxes.push([...newBox.yxBox]);
-        }
       }
     }
-    resolve(bodies);
+    resolve(cache.bodies);
   });
 }

src/body/movenetcoords.ts

@@ -1,4 +1,4 @@
-export const kpt = [
+export const kpt: Array<string> = [
   'nose',
   'leftEye',
   'rightEye',
@@ -18,7 +18,7 @@ export const kpt = [
   'rightAnkle',
 ];
 
-export const connected = {
+export const connected: Record<string, string[]> = {
   leftLeg: ['leftHip', 'leftKnee', 'leftAnkle'],
   rightLeg: ['rightHip', 'rightKnee', 'rightAnkle'],
   torso: ['leftShoulder', 'rightShoulder', 'rightHip', 'leftHip', 'leftShoulder'],

src/config.ts

@@ -420,12 +420,12 @@ const config: Config = {
     rotation: true,          // use best-guess rotated hand image or just box with rotation as-is
                              // false means higher performance, but incorrect finger mapping if hand is inverted
                              // only valid for `handdetect` variation
-    skipFrames: 14,          // how many max frames to go without re-running the hand bounding box detector
+    skipFrames: 1,           // how many max frames to go without re-running the hand bounding box detector
                              // only used when cacheSensitivity is not zero
                              // e.g., if model is running st 25 FPS, we can re-use existing bounding
                              // box for updated hand skeleton analysis as the hand
                              // hasn't moved much in short time (10 * 1/25 = 0.25 sec)
-    minConfidence: 0.5,      // threshold for discarding a prediction
+    minConfidence: 0.55,     // threshold for discarding a prediction
     iouThreshold: 0.2,       // ammount of overlap between two detected objects before one object is removed
     maxDetected: -1,         // maximum number of hands detected in the input
                              // should be set to the minimum number for performance

src/hand/handtrack.ts

@@ -7,7 +7,7 @@
  */
 
 import { log, join } from '../util/util';
-import { scale } from '../util/box';
+import * as box from '../util/box';
 import * as tf from '../../dist/tfjs.esm.js';
 import type { HandResult, Box, Point } from '../result';
 import type { GraphModel, Tensor } from '../tfjs/types';
@@ -16,7 +16,6 @@ import { env } from '../util/env';
 import * as fingerPose from './fingerpose';
 import { fakeOps } from '../tfjs/backend';
 
-const boxScaleFact = 1.5; // hand finger model prefers slighly larger box
 const models: [GraphModel | null, GraphModel | null] = [null, null];
 const modelOutputNodes = ['StatefulPartitionedCall/Postprocessor/Slice', 'StatefulPartitionedCall/Postprocessor/ExpandDims_1'];
 
@@ -24,26 +23,26 @@ const inputSize = [[0, 0], [0, 0]];
 
 const classes = ['hand', 'fist', 'pinch', 'point', 'face', 'tip', 'pinchtip'];
 
+const boxExpandFact = 1.6; // increase to 160%
+
 let skipped = 0;
-let outputSize: Point = [0, 0];
+let outputSize: [number, number] = [0, 0];
 
 type HandDetectResult = {
   id: number,
   score: number,
   box: Box,
   boxRaw: Box,
+  boxCrop: Box,
   label: string,
-  yxBox: Box,
 }
 
 const cache: {
-  handBoxes: Array<HandDetectResult>,
+  boxes: Array<HandDetectResult>,
-  fingerBoxes: Array<HandDetectResult>
+  hands: Array<HandResult>;
-  tmpBoxes: Array<HandDetectResult>
 } = {
-  handBoxes: [],
+  boxes: [],
-  fingerBoxes: [],
+  hands: [],
-  tmpBoxes: [],
 };
 
 const fingerMap = {
@@ -103,35 +102,29 @@ async function detectHands(input: Tensor, config: Config): Promise<HandDetectRes
   [t.rawScores, t.rawBoxes] = await models[0].executeAsync(t.cast, modelOutputNodes) as Tensor[];
   t.boxes = tf.squeeze(t.rawBoxes, [0, 2]);
   t.scores = tf.squeeze(t.rawScores, [0]);
-  const classScores = tf.unstack(t.scores, 1);
+  const classScores = tf.unstack(t.scores, 1); // unstack scores based on classes
+  classScores.splice(4, 1); // remove faces
+  t.filtered = tf.stack(classScores, 1); // restack
+  tf.dispose(...classScores);
+  t.max = tf.max(t.filtered, 1); // max overall score
+  t.argmax = tf.argMax(t.filtered, 1); // class index of max overall score
   let id = 0;
-  for (let i = 0; i < classScores.length; i++) {
+  t.nms = await tf.image.nonMaxSuppressionAsync(t.boxes, t.max, config.hand.maxDetected, config.hand.iouThreshold, config.hand.minConfidence);
-    if (i === 4) continue; // skip faces
+  const nms = await t.nms.data();
-    t.nms = await tf.image.nonMaxSuppressionAsync(t.boxes, classScores[i], config.hand.maxDetected, config.hand.iouThreshold, config.hand.minConfidence);
+  const scores = await t.max.data();
-    const nms = await t.nms.data();
+  const classNum = await t.argmax.data();
-    tf.dispose(t.nms);
+  for (const nmsIndex of Array.from(nms)) { // generates results for each class
-    for (const res of Array.from(nms)) { // generates results for each class
+    const boxSlice = tf.slice(t.boxes, nmsIndex, 1);
-      const boxSlice = tf.slice(t.boxes, res, 1);
+    const boxData = await boxSlice.data();
-      let yxBox: Box = [0, 0, 0, 0];
+    tf.dispose(boxSlice);
-      if (config.hand.landmarks) { // scale box
+    const boxInput: Box = [boxData[1], boxData[0], boxData[3] - boxData[1], boxData[2] - boxData[0]];
-        const detectedBox: Box = await boxSlice.data();
+    const boxRaw: Box = box.scale(boxInput, 1.2); // handtrack model returns tight box so we expand it a bit
-        const boxCenter: Point = [(detectedBox[0] + detectedBox[2]) / 2, (detectedBox[1] + detectedBox[3]) / 2];
+    const boxFull: Box = [Math.trunc(boxRaw[0] * outputSize[0]), Math.trunc(boxRaw[1] * outputSize[1]), Math.trunc(boxRaw[2] * outputSize[0]), Math.trunc(boxRaw[3] * outputSize[1])];
-        const boxDiff: Box = [+boxCenter[0] - detectedBox[0], +boxCenter[1] - detectedBox[1], -boxCenter[0] + detectedBox[2], -boxCenter[1] + detectedBox[3]];
+    const score = scores[nmsIndex];
-        yxBox = [boxCenter[0] - boxScaleFact * boxDiff[0], boxCenter[1] - boxScaleFact * boxDiff[1], boxCenter[0] + boxScaleFact * boxDiff[2], boxCenter[1] + boxScaleFact * boxDiff[3]];
+    const label = classes[classNum[nmsIndex]];
-      } else { // use box as-is
+    const hand: HandDetectResult = { id: id++, score, box: boxFull, boxRaw, boxCrop: box.crop(boxRaw), label };
-        yxBox = await boxSlice.data();
+    hands.push(hand);
-      }
-      const boxRaw: Box = [yxBox[1], yxBox[0], yxBox[3] - yxBox[1], yxBox[2] - yxBox[0]];
-      const box: Box = [Math.trunc(boxRaw[0] * outputSize[0]), Math.trunc(boxRaw[1] * outputSize[1]), Math.trunc(boxRaw[2] * outputSize[0]), Math.trunc(boxRaw[3] * outputSize[1])];
-      tf.dispose(boxSlice);
-      const scoreSlice = tf.slice(classScores[i], res, 1);
-      const score = (await scoreSlice.data())[0];
-      tf.dispose(scoreSlice);
-      const hand: HandDetectResult = { id: id++, score, box, boxRaw, label: classes[i], yxBox };
-      hands.push(hand);
-    }
   }
-  classScores.forEach((tensor) => tf.dispose(tensor));
   Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));
   hands.sort((a, b) => b.score - a.score);
   if (hands.length > (config.hand.maxDetected || 1)) hands.length = (config.hand.maxDetected || 1);
@@ -139,7 +132,7 @@ async function detectHands(input: Tensor, config: Config): Promise<HandDetectRes
 }
 
 async function detectFingers(input: Tensor, h: HandDetectResult, config: Config): Promise<HandResult> {
-  const hand: HandResult = {
+  const hand: HandResult = { // initial values inherited from hand detect
     id: h.id,
     score: Math.round(100 * h.score) / 100,
     boxScore: Math.round(100 * h.score) / 100,
@@ -151,36 +144,27 @@ async function detectFingers(input: Tensor, h: HandDetectResult, config: Config)
     landmarks: {} as HandResult['landmarks'],
     annotations: {} as HandResult['annotations'],
   };
-  if (input && models[1] && config.hand.landmarks) {
+  if (input && models[1] && config.hand.landmarks && h.score > (config.hand.minConfidence || 0)) {
     const t: Record<string, Tensor> = {};
-    if (!h.yxBox) return hand;
+    t.crop = tf.image.cropAndResize(input, [box.crop(h.boxRaw)], [0], [inputSize[1][0], inputSize[1][1]], 'bilinear');
-    t.crop = tf.image.cropAndResize(input, [h.yxBox], [0], [inputSize[1][0], inputSize[1][1]], 'bilinear');
     t.cast = tf.cast(t.crop, 'float32');
     t.div = tf.div(t.cast, 255);
     [t.score, t.keypoints] = models[1].execute(t.div) as Tensor[];
-    // const score = Math.round(100 * (await t.score.data())[0] / 100);
     const rawScore = (await t.score.data())[0];
     const score = (100 - Math.trunc(100 / (1 + Math.exp(rawScore)))) / 100; // reverse sigmoid value
     if (score >= (config.hand.minConfidence || 0)) {
       hand.fingerScore = score;
       t.reshaped = tf.reshape(t.keypoints, [-1, 3]);
       const rawCoords = await t.reshaped.array() as Point[];
-      hand.keypoints = (rawCoords as Point[]).map((coord) => [
+      hand.keypoints = (rawCoords as Point[]).map((kpt) => [
-        (h.box[2] * coord[0] / inputSize[1][0]) + h.box[0],
+        outputSize[0] * ((h.boxCrop[3] - h.boxCrop[1]) * kpt[0] / inputSize[1][0] + h.boxCrop[1]),
-        (h.box[3] * coord[1] / inputSize[1][1]) + h.box[1],
+        outputSize[1] * ((h.boxCrop[2] - h.boxCrop[0]) * kpt[1] / inputSize[1][1] + h.boxCrop[0]),
-        (h.box[2] + h.box[3]) / 2 / inputSize[1][0] * (coord[2] || 0),
+        (h.boxCrop[3] + h.boxCrop[3] / 2 * (kpt[2] || 0)),
       ]);
-      const updatedBox = scale(hand.keypoints, boxScaleFact, outputSize); // replace detected box with box calculated around keypoints
-      h.box = updatedBox.box;
-      h.boxRaw = updatedBox.boxRaw;
-      h.yxBox = updatedBox.yxBox;
-      hand.box = h.box;
       hand.landmarks = fingerPose.analyze(hand.keypoints) as HandResult['landmarks']; // calculate finger landmarks
       for (const key of Object.keys(fingerMap)) { // map keypoints to per-finger annotations
         hand.annotations[key] = fingerMap[key].map((index) => (hand.landmarks && hand.keypoints[index] ? hand.keypoints[index] : null));
       }
-      const ratioBoxFrame = Math.min(h.box[2] / (input.shape[2] || 1), h.box[3] / (input.shape[1] || 1));
-      if (ratioBoxFrame > 0.05) cache.tmpBoxes.push(h); // if finger detection is enabled, only update cache if fingers are detected and box is big enough
     }
     Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));
   }
@@ -188,22 +172,37 @@ async function detectFingers(input: Tensor, h: HandDetectResult, config: Config)
 }
 
 export async function predict(input: Tensor, config: Config): Promise<HandResult[]> {
+  if (!models[0] || !models[1] || !models[0]?.inputs[0].shape || !models[1]?.inputs[0].shape) return []; // something is wrong with the model
   outputSize = [input.shape[2] || 0, input.shape[1] || 0];
-  let hands: Array<HandResult> = [];
+
-  cache.tmpBoxes = []; // clear temp cache
+  skipped++; // increment skip frames
-  if (!config.hand.landmarks) cache.fingerBoxes = cache.handBoxes; // if hand detection only reset finger boxes cache
+  if (config.skipFrame && (skipped <= (config.hand.skipFrames || 0))) {
-  if (!config.skipFrame) cache.fingerBoxes = [];
+    return cache.hands; // return cached results without running anything
-  if ((skipped < (config.hand.skipFrames || 0)) && config.skipFrame) { // just run finger detection while reusing cached boxes
-    skipped++;
-    hands = await Promise.all(cache.fingerBoxes.map((hand) => detectFingers(input, hand, config))); // run from finger box cache
-  } else { // calculate new boxes and run finger detection
-    skipped = 0;
-    hands = await Promise.all(cache.fingerBoxes.map((hand) => detectFingers(input, hand, config))); // run from finger box cache
-    if (hands.length !== config.hand.maxDetected) { // re-run with hand detection only if we dont have enough hands in cache
-      cache.handBoxes = await detectHands(input, config);
-      hands = await Promise.all(cache.handBoxes.map((hand) => detectFingers(input, hand, config)));
-    }
   }
-  cache.fingerBoxes = [...cache.tmpBoxes]; // repopulate cache with validated hands
+  return new Promise(async (resolve) => {
-  return hands as HandResult[];
+    skipped = 0;
+    if (cache.boxes.length >= (config.hand.maxDetected || 0)) {
+      cache.hands = await Promise.all(cache.boxes.map((handBox) => detectFingers(input, handBox, config))); // if we have enough cached boxes run detection using cache
+    } else {
+      cache.hands = []; // reset hands
+    }
+
+    if (cache.hands.length !== config.hand.maxDetected) { // did not find enough hands based on cached boxes so run detection on full frame
+      cache.boxes = await detectHands(input, config);
+      cache.hands = await Promise.all(cache.boxes.map((handBox) => detectFingers(input, handBox, config)));
+    }
+
+    const oldCache = [...cache.boxes];
+    cache.boxes.length = 0; // reset cache
+    for (let i = 0; i < cache.hands.length; i++) {
+      const boxKpt = box.square(cache.hands[i].keypoints, outputSize);
+      if (boxKpt.box[2] / (input.shape[2] || 1) > 0.05 && boxKpt.box[3] / (input.shape[1] || 1) > 0.05 && cache.hands[i].fingerScore && cache.hands[i].fingerScore > (config.hand.minConfidence || 0)) {
+        const boxScale = box.scale(boxKpt.box, boxExpandFact);
+        const boxScaleRaw = box.scale(boxKpt.boxRaw, boxExpandFact);
+        const boxCrop = box.crop(boxScaleRaw);
+        cache.boxes.push({ ...oldCache[i], box: boxScale, boxRaw: boxScaleRaw, boxCrop });
+      }
+    }
+    resolve(cache.hands);
+  });
 }

src/human.ts

@@ -359,7 +359,7 @@ export class Human {
    * @returns result: {@link Result}
    */
   next(result: Result = this.result): Result {
-    return interpolate.calc(result) as Result;
+    return interpolate.calc(result, this.config) as Result;
   }
 
   /** Warmup method pre-initializes all configured models for faster inference

src/result.ts

@@ -84,7 +84,7 @@ export interface BodyResult {
   score: number,
   box: Box,
   boxRaw: Box,
-  annotations: Record<string, Point[][]>,
+  annotations: Record<string, Array<Point[]>>,
   keypoints: Array<BodyKeypoint>
 }
 

src/util/box.ts

@@ -1,28 +1,32 @@
-import type { Box } from '../result';
+import type { Point, Box } from '../result';
 
-// helper function: find box around keypoints, square it and scale it
+export function calc(keypoints: Array<Point>, outputSize: [number, number] = [1, 1]) {
-export function scale(keypoints, boxScaleFact, outputSize) {
   const coords = [keypoints.map((pt) => pt[0]), keypoints.map((pt) => pt[1])]; // all x/y coords
-  const maxmin = [Math.max(...coords[0]), Math.min(...coords[0]), Math.max(...coords[1]), Math.min(...coords[1])]; // find min/max x/y coordinates
+  const min = [Math.min(...coords[0]), Math.min(...coords[1])];
-  const center = [(maxmin[0] + maxmin[1]) / 2, (maxmin[2] + maxmin[3]) / 2]; // find center x and y coord of all fingers
+  const max = [Math.max(...coords[0]), Math.max(...coords[1])];
-  const diff = Math.max(center[0] - maxmin[1], center[1] - maxmin[3], -center[0] + maxmin[0], -center[1] + maxmin[2]) * boxScaleFact; // largest distance from center in any direction
+  const box: Box = [min[0], min[1], max[0] - min[0], max[1] - min[1]];
-  const box = [
+  const boxRaw: Box = [box[0] / outputSize[0], box[1] / outputSize[1], box[2] / outputSize[0], box[3] / outputSize[1]];
-    Math.trunc(center[0] - diff),
+  return { box, boxRaw };
-    Math.trunc(center[1] - diff),
+}
-    Math.trunc(2 * diff),
+
-    Math.trunc(2 * diff),
+export function square(keypoints: Array<Point>, outputSize: [number, number] = [1, 1]) {
-  ] as Box;
+  const coords = [keypoints.map((pt) => pt[0]), keypoints.map((pt) => pt[1])]; // all x/y coords
-  const boxRaw = [ // work backwards
+  const min = [Math.min(...coords[0]), Math.min(...coords[1])];
-    box[0] / outputSize[0],
+  const max = [Math.max(...coords[0]), Math.max(...coords[1])];
-    box[1] / outputSize[1],
+  const center = [(min[0] + max[0]) / 2, (min[1] + max[1]) / 2]; // find center x and y coord of all fingers
-    box[2] / outputSize[0],
+  const dist = Math.max(center[0] - min[0], center[1] - min[1], -center[0] + max[0], -center[1] + max[1]); // largest distance from center in any direction
-    box[3] / outputSize[1],
+  const box: Box = [Math.trunc(center[0] - dist), Math.trunc(center[1] - dist), Math.trunc(2 * dist), Math.trunc(2 * dist)];
-  ] as Box;
+  const boxRaw: Box = [box[0] / outputSize[0], box[1] / outputSize[1], box[2] / outputSize[0], box[3] / outputSize[1]];
-  const yxBox = [ // work backwards
+  return { box, boxRaw };
-    boxRaw[1],
+}
-    boxRaw[0],
+
-    boxRaw[3] + boxRaw[1],
+export function scale(box: Box, scaleFact: number) {
-    boxRaw[2] + boxRaw[0],
+  const dist = [box[2] * (scaleFact - 1), box[3] * (scaleFact - 1)];
-  ] as Box;
+  const newBox: Box = [box[0] - dist[0] / 2, box[1] - dist[1] / 2, box[2] + dist[0], box[3] + dist[0]];
-  return { box, boxRaw, yxBox };
+  return newBox;
+}
+
+export function crop(box: Box) { // [y1, x1, y2, x2] clamped to 0..1
+  const yxBox: Box = [Math.max(0, box[1]), Math.max(0, box[0]), Math.min(1, box[3] + box[1]), Math.min(1, box[2] + box[0])];
+  return yxBox;
 }

src/util/interpolate.ts

@@ -3,10 +3,16 @@
  */
 
 import type { Result, FaceResult, BodyResult, HandResult, ObjectResult, GestureResult, PersonResult, Box, Point } from '../result';
+import type { Config } from '../config';
+
+import * as moveNetCoords from '../body/movenetcoords';
+import * as blazePoseCoords from '../body/blazeposecoords';
+import * as efficientPoseCoords from '../body/efficientposecoords';
 
 const bufferedResult: Result = { face: [], body: [], hand: [], gesture: [], object: [], persons: [], performance: {}, timestamp: 0 };
 
-export function calc(newResult: Result): Result {
+export function calc(newResult: Result, config: Config): Result {
+  const t0 = performance.now();
   if (!newResult) return { face: [], body: [], hand: [], gesture: [], object: [], persons: [], performance: {}, timestamp: 0 };
   // each record is only updated using deep clone when number of detected record changes, otherwise it will converge by itself
   // otherwise bufferedResult is a shallow clone of result plus updated local calculated values
@@ -46,7 +52,22 @@ export function calc(newResult: Result): Result {
             bufferedResult.body[i].keypoints[j] ? ((bufferedFactor - 1) * bufferedResult.body[i].keypoints[j].positionRaw[1] + keypoint.positionRaw[1]) / bufferedFactor : keypoint.position[1],
           ],
         }))) as Array<{ score: number, part: string, position: [number, number, number?], positionRaw: [number, number, number?] }>;
-      bufferedResult.body[i] = { ...newResult.body[i], box, boxRaw, keypoints }; // shallow clone plus updated values
+      const annotations: Record<string, Point[][]> = {};
+
+      let coords = { connected: {} };
+      if (config.body?.modelPath?.includes('efficientpose')) coords = efficientPoseCoords;
+      else if (config.body?.modelPath?.includes('blazepose')) coords = blazePoseCoords;
+      else if (config.body?.modelPath?.includes('movenet')) coords = moveNetCoords;
+      for (const [name, indexes] of Object.entries(coords.connected as Record<string, string[]>)) {
+        const pt: Array<Point[]> = [];
+        for (let j = 0; j < indexes.length - 1; j++) {
+          const pt0 = keypoints.find((kp) => kp.part === indexes[j]);
+          const pt1 = keypoints.find((kp) => kp.part === indexes[j + 1]);
+          if (pt0 && pt1 && pt0.score > (config.body.minConfidence || 0) && pt1.score > (config.body.minConfidence || 0)) pt.push([pt0.position, pt1.position]);
+        }
+        annotations[name] = pt;
+      }
+      bufferedResult.body[i] = { ...newResult.body[i], box, boxRaw, keypoints, annotations: annotations as BodyResult['annotations'] }; // shallow clone plus updated values
     }
   }
 
@@ -64,12 +85,16 @@ export function calc(newResult: Result): Result {
         .map((landmark, j) => landmark
           .map((coord, k) => (((bufferedFactor - 1) * (bufferedResult.hand[i].keypoints[j][k] || 1) + (coord || 0)) / bufferedFactor)) as Point)
         : [];
-      const annotations = {};
+      let annotations = {};
-      if (Object.keys(bufferedResult.hand[i].annotations).length !== Object.keys(newResult.hand[i].annotations).length) bufferedResult.hand[i].annotations = newResult.hand[i].annotations; // reset annotations as previous frame did not have them
+      if (Object.keys(bufferedResult.hand[i].annotations).length !== Object.keys(newResult.hand[i].annotations).length) {
-      if (newResult.hand[i].annotations) {
+        bufferedResult.hand[i].annotations = newResult.hand[i].annotations; // reset annotations as previous frame did not have them
+        annotations = bufferedResult.hand[i].annotations;
+      } else if (newResult.hand[i].annotations) {
         for (const key of Object.keys(newResult.hand[i].annotations)) { // update annotations
           annotations[key] = newResult.hand[i].annotations[key] && newResult.hand[i].annotations[key][0]
-            ? newResult.hand[i].annotations[key].map((val, j) => val.map((coord, k) => ((bufferedFactor - 1) * bufferedResult.hand[i].annotations[key][j][k] + coord) / bufferedFactor))
+            ? newResult.hand[i].annotations[key]
+              .map((val, j) => val
+                .map((coord, k) => ((bufferedFactor - 1) * bufferedResult.hand[i].annotations[key][j][k] + coord) / bufferedFactor))
             : null;
         }
       }
@@ -134,7 +159,10 @@ export function calc(newResult: Result): Result {
 
   // just copy latest gestures without interpolation
   if (newResult.gesture) bufferedResult.gesture = newResult.gesture as GestureResult[];
-  if (newResult.performance) bufferedResult.performance = newResult.performance;
+
+  // append interpolation performance data
+  const t1 = performance.now();
+  if (newResult.performance) bufferedResult.performance = { ...newResult.performance, interpolate: Math.round(t1 - t0) };
 
   return bufferedResult;
 }