This repository has no description
feat(android): higher pose input res in CROP + conf filter + smoothing
Three compounding pose-quality tweaks:
- CROP downscale target rises from 256 → 384 max side. The crop is
already smaller than the full frame so this lands more pixels on the
person without moving the mask-path downscale.
- Drop landmarks whose MLKit inFrameLikelihood is below 0.5 before
emitting. Removes the "phantom joint" problem where occluded limbs
get guessed with low confidence.
- Temporal smoothing on successive skeletons (α=0.6, 500ms gap reset)
via PoseSmoother. Cuts visible jitter during fast shot motion while
staying responsive to real position changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
author
virtualintern
co-author
Claude Opus 4.7 (1M context)
date
(Apr 20, 2026, 6:34 PM +0300)
commit
da0d404c
da0d404c9d1a3bf5bd805219ff1596572217e0b7
parent
b7c63078
b7c63078b2c27c10084c9121f15fbae5d3a04119
+43
-3
+43
-3
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/camera/Utils.android.kt
+43
-3
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/camera/Utils.android.kt
···
248
248
// 2) Pose input: build MLKit image (optionally masked) and DOWNscale it.
249
249
// Rotation is now 0 because analysisBitmap is upright.
250
250
val mlKitPoseInput: MlKitPoseInput? = poseDetector?.let {
251
+
// In CROP mode we feed MLKit only the focus-area crop, which is already
252
+
// smaller than the full frame — so raise the downscale target to give
253
+
// the model more effective pixels on the person without ballooning work.
254
+
val downscale = if (focusArea != null && poseFocusMode == PoseFocusMode.CROP) 384 else 256
251
255
buildMlKitPoseInput(
252
256
analysisBitmap = analysisBitmap,
253
257
focusArea = focusArea,
254
258
poseFocusMode = poseFocusMode,
255
-
downscaleMaxSidePx = 256
259
+
downscaleMaxSidePx = downscale
256
260
)
257
261
}
258
262
···
292
296
// Shared executor for running pose detection in parallel with object detection.
293
297
private val poseExecutor = Executors.newSingleThreadExecutor()
294
298
299
+
// Drop landmarks whose MLKit inFrameLikelihood is below this threshold — keeps
300
+
// obvious phantoms (occluded joints guessed with low conf) out of downstream.
301
+
private const val LANDMARK_CONF_THRESHOLD = 0.5f
302
+
303
+
// Temporal smoothing across consecutive skeletons reduces jitter during fast
304
+
// shot motion. α is the weight on the new frame; (1-α) on the previous.
305
+
private const val SMOOTHING_ALPHA = 0.6f
306
+
// Reset the smoother after a long gap so a re-acquired pose doesn't inherit a
307
+
// stale position from where the shooter used to be.
308
+
private const val SMOOTHING_GAP_MS = 500L
309
+
310
+
private object PoseSmoother {
311
+
private var last: Skeleton? = null
312
+
private var lastWallMs: Long = 0L
313
+
314
+
@Synchronized
315
+
fun smooth(skel: Skeleton?, timestamp: Long): Skeleton? {
316
+
if (skel == null) {
317
+
if (timestamp - lastWallMs > SMOOTHING_GAP_MS) last = null
318
+
return null
319
+
}
320
+
val prev = last
321
+
val gapOk = (timestamp - lastWallMs) <= SMOOTHING_GAP_MS
322
+
val out = if (prev != null && gapOk) Skeleton.lerp(prev, skel, SMOOTHING_ALPHA) else skel
323
+
last = out
324
+
lastWallMs = timestamp
325
+
return out
326
+
}
327
+
}
328
+
295
329
// Core processing used by both live camera (ImageProxy) and offline frames (FrameAnalyser).
296
330
fun process(
297
331
tensorImage: TensorImage?,
···
313
347
poseExecutor.submit<Skeleton?> {
314
348
runCatching {
315
349
val pose = Tasks.await(poseDetector.process(mlKitImage))
316
-
skeletonFromPoseScaled(
350
+
val raw = skeletonFromPoseScaled(
317
351
pose = pose,
318
352
timestamp = timestamp,
319
353
width = width,
···
323
357
offsetX = mlKitOffsetX,
324
358
offsetY = mlKitOffsetY,
325
359
)
360
+
// Empty skeletons (no landmarks passed conf threshold) smooth as
361
+
// null so the smoother doesn't lerp toward a zeroed pose.
362
+
val hasAny = raw.joints().isNotEmpty()
363
+
PoseSmoother.smooth(if (hasAny) raw else null, timestamp)
326
364
}.onFailure { t ->
327
365
Logger.e(t) { "MLKit poseDetector.process failed" }
328
366
}.getOrNull()
···
816
854
offsetX: Float,
817
855
offsetY: Float,
818
856
): Skeleton.SkeletonCoordinate? {
819
-
val pos = this?.position ?: return null
857
+
val lm = this ?: return null
858
+
if (lm.inFrameLikelihood < LANDMARK_CONF_THRESHOLD) return null
859
+
val pos = lm.position
820
860
return Skeleton.SkeletonCoordinate(
821
861
x = pos.x * scaleX + offsetX,
822
862
y = pos.y * scaleY + offsetY,