Don't detrend for long videos; Add warning for simple rPPG methods

vitallens/client.py

@@ -45,7 +45,7 @@ def __init__(
       method: Method = Method.VITALLENS,
       api_key: str = None,
       detect_faces: bool = True,
-      fdet_max_faces: int = 2,
+      fdet_max_faces: int = 1,
       fdet_fs: float = 1.0,
       fdet_score_threshold: float = 0.9,
       fdet_iou_threshold: float = 0.3
@@ -65,6 +65,7 @@ def __init__(
     self.api_key = api_key
     # Load the config and model
     self.config = load_config(method.name.lower() + ".yaml")
+    self.method = method
     if self.config['model'] == 'g':
       self.rppg = GRPPGMethod(self.config)
     elif self.config['model'] == 'chrom':
@@ -151,6 +152,11 @@ def __call__(
     """
     # Probe inputs
     inputs_shape, fps = probe_video_inputs(video=video, fps=fps)
+    # TODO: Optimize performance of simple rPPG methods for long videos
+    # Warning if using long video
+    target_fps = override_fps_target if override_fps_target is not None else self.rppg.fps_target
+    if self.method != Method.VITALLENS and inputs_shape[0]/fps*target_fps > 3600:
+      logging.warn("Inference for long videos has yet to be optimized for POS / G / CHROM. This may run out of memory and crash.")
     _, height, width, _ = inputs_shape
     if self.detect_faces:
       # Detect faces

vitallens/methods/vitallens.py

@@ -79,6 +79,7 @@ def __call__(
       (faces[:,3] - faces[:,1]) * 0.5 < np.maximum(0, faces[:,1] - roi[1]) + np.maximum(0, faces[:,3] - roi[3]))):
       logging.warn("Large face movement detected")
     # Parse the inputs
+    logging.debug("Preparing video for inference...")
     frames_ds, fps, inputs_shape, ds_factor, _ = parse_video_inputs(
       video=frames, fps=fps, target_size=self.input_size, roi=roi,
       target_fps=override_fps_target if override_fps_target is not None else self.fps_target,
@@ -95,6 +96,7 @@ def __call__(
       split_len = math.ceil((ds_len + (n_splits-1) * API_OVERLAP) / n_splits)
       start_idxs = [i for i in range(0, ds_len - n_splits * API_OVERLAP, split_len - API_OVERLAP)]
       end_idxs = [min(i + split_len, ds_len) for i in start_idxs]
+      logging.info("Running inference for {} frames using {} requests...".format(ds_len, n_splits))
       # Process the splits in parallel
       with concurrent.futures.ThreadPoolExecutor() as executor:
         results = list(executor.map(lambda i: self.process_api(frames_ds[start_idxs[i]:end_idxs[i]]), range(n_splits)))
@@ -211,8 +213,9 @@ def postprocess(self, sig, fps, type='ppg', filter=True):
         Lambda = detrend_lambda_for_rr_response(fps)
       else:
         raise ValueError("Type {} not implemented!".format(type))
-      # Detrend
-      sig = detrend(sig, Lambda)
+      if sig.shape[-1] < 4 * API_MAX_FRAMES:
+        # Detrend only for shorter videos for performance reasons
+        sig = detrend(sig, Lambda)
       # Moving average
       sig = moving_average(sig, size)
       # Standardize