Merge pull request #1 from sovrasov/asynch_detector_call

Asynch detector call

Author: DmitriySidnev

demos/python_demos/multi_camera_multi_person_tracking/multi_camera_multi_person_tracking.py

@@ -80,6 +80,9 @@ def run(params, capture, detector, reid):
     else:
         output_video = None
 
+    prev_frames = thread_body.frames_queue.get()
+    detector.run_asynch(prev_frames, frame_number)
+
     while thread_body.process:
         key = check_pressed_keys(key)
         if key == 27:
@@ -93,30 +96,30 @@ def run(params, capture, detector, reid):
         if frames is None:
             continue
 
-        if params.detections:
-            all_detections = detector.get_detections(frame_number)
-        else:
-            all_detections = detector.get_detections(frames)
+        frame_number += 1
+        all_detections = detector.wait_and_grab()
+        detector.run_asynch(frames, frame_number)
+
         all_masks = [[] for _ in range(len(all_detections))]
         for i, detections in enumerate(all_detections):
             all_detections[i] = [det[0] for det in detections]
             all_masks[i] = [det[2] for det in detections if len(det) == 3]
 
-        tracker.process(frames, all_detections, all_masks)
+        tracker.process(prev_frames, all_detections, all_masks)
         tracked_objects = tracker.get_tracked_objects()
 
         latency = time.time() - start
         avg_latency.update(latency)
         fps = round(1. / latency, 1)
 
-        vis = visualize_multicam_detections(frames, tracked_objects, fps, **config['visualization_config'])
+        vis = visualize_multicam_detections(prev_frames, tracked_objects, fps, **config['visualization_config'])
         cv.imshow(win_name, vis)
         if output_video:
             output_video.write(cv.resize(vis, video_output_size))
 
         print('\rProcessing frame: {}, fps = {} (avg_fps = {:.3})'.format(
                             frame_number, fps, 1. / avg_latency.get()), end="")
-        frame_number += 1
+        prev_frames, frames = frames, prev_frames
     print('')
 
     thread_body.process = False

demos/python_demos/multi_camera_multi_person_tracking/utils/network_wrappers.py

@@ -14,6 +14,7 @@
 import json
 import logging as log
 from collections import namedtuple
+from abc import ABC, abstractmethod
 
 import cv2
 import numpy as np
@@ -22,7 +23,17 @@
 from .segm_postrocess import postprocess
 
 
-class Detector:
+class DetectorInterface(ABC):
+    @abstractmethod
+    def run_asynch(self, frames, index):
+        pass
+
+    @abstractmethod
+    def wait_and_grab(self):
+        pass
+
+
+class Detector(DetectorInterface):
     """Wrapper class for detector"""
 
     def __init__(self, ie, model_path, conf=.6, device='CPU', ext_path='', max_num_frames=1):
@@ -31,21 +42,26 @@ def __init__(self, ie, model_path, conf=.6, device='CPU', ext_path='', max_num_f
         self.expand_ratio = (1., 1.)
         self.max_num_frames = max_num_frames
 
-    def get_detections(self, frames):
-        """Returns all detections on frames"""
+    def run_asynch(self, frames, index):
         assert len(frames) <= self.max_num_frames
-
-        all_detections = []
+        self.shapes = []
         for i in range(len(frames)):
+            self.shapes.append(frames[i].shape)
             self.net.forward_async(frames[i])
-        outputs = self.net.grab_all_async()
 
+    def wait_and_grab(self):
+        all_detections = []
+        outputs = self.net.grab_all_async()
         for i, out in enumerate(outputs):
-            detections = self.__decode_detections(out, frames[i].shape)
+            detections = self.__decode_detections(out, self.shapes[i])
             all_detections.append(detections)
-
         return all_detections
 
+    def get_detections(self, frames):
+        """Returns all detections on frames"""
+        self.run_asynch(frames)
+        return self.wait_and_grab()
+
     def __decode_detections(self, out, frame_shape):
         """Decodes raw SSD output"""
         detections = []
@@ -130,7 +146,7 @@ def forward(self, rois):
         return embeddings
 
 
-class MaskRCNN:
+class MaskRCNN(DetectorInterface):
     """Wrapper class for a network returning masks of objects"""
 
     def __init__(self, ie, model_path, conf=.6, device='CPU', ext_path='',
@@ -213,6 +229,12 @@ def get_detections(self, frames, return_cropped_masks=True, only_class_person=Tr
             outputs.append(frame_output)
         return outputs
 
+    def run_asynch(self, frames, index):
+        self.frames = frames
+
+    def wait_and_grab(self):
+        return self.get_detections(self.frames)
+
     class Compose(object):
         def __init__(self, transforms):
             self.transforms = transforms
@@ -286,7 +308,7 @@ def __call__(self, sample):
             return sample
 
 
-class DetectionsFromFileReader(object):
+class DetectionsFromFileReader(DetectorInterface):
     """Read detection from *.json file.
     Format of the file should be:
     [
@@ -296,6 +318,7 @@ class DetectionsFromFileReader(object):
         ...
     ]
     """
+
     def __init__(self, input_files, score_thresh):
         self.input_files = input_files
         self.score_thresh = score_thresh
@@ -309,12 +332,15 @@ def __init__(self, input_files, score_thresh):
                     detections_dict[det['frame_id']] = {'boxes': det['boxes'], 'scores': det['scores']}
                 self.detections.append(detections_dict)
 
-    def get_detections(self, frame_id):
+    def run_asynch(self, frames, index):
+        self.last_index = index
+
+    def wait_and_grab(self):
         output = []
         for source in self.detections:
             valid_detections = []
-            if frame_id in source:
-                for bbox, score in zip(source[frame_id]['boxes'], source[frame_id]['scores']):
+            if self.last_index in source:
+                for bbox, score in zip(source[self.last_index]['boxes'], source[self.last_index]['scores']):
                     if score > self.score_thresh:
                         bbox = [int(value) for value in bbox]
                         valid_detections.append((bbox, score))