Robust association, optimization, and API restructure (#175)

* Add robust feature

* Speed up cdist with Numba

* Handle occluded tracks

* Remove clipping in detector

* Fix smooth dist

* Fix cluster distance

* Add assertion for public detector

* Account for negative width or height for area

* Add confirm_hits parameter

* Speed up numpy concat

* Fix ReID

* Normalize cluster feature

* Speed up clustering using cosine dist

* clean up

* Refactor config parameter parsing

* Fix letterbox and add parameter doc

* Use greedy match for ReID and improve remove dup

* Split into multiple lines

* Experiment with match rectification

* Speed up rect and bbox distance helper functions

* Speed up ios and iom

* Deprecate cython-bbox

* Speed up matching and gating

* Speed up rect and fix cdist

* Threshold aspect ratio and more optimizations

* Allow defining model class outside the API

* Add comments

* Speed up matching cascade and sync with tensorrt_demos

* Add doc

* Add visualizer parameters

* Lower fill_val and clean up

* Clean up

Dockerfile

@@ -83,7 +83,6 @@ ARG CUPY_NVCC_GENERATE_CODE
 RUN if [[ -z ${CUPY_NVCC_GENERATE_CODE} ]]; then \
         echo "CUPY_NVCC_GENERATE_CODE not set, building CuPy for all architectures (slower)"; \
     fi && \
-    pip install --no-cache-dir cython && \
     if [[ ${TRT_IMAGE_VERSION} == 21.05 ]]; then \
         CUPY_NUM_BUILD_JOBS=$(nproc) pip install --no-cache-dir -r <(grep -ivE "tensorflow" requirements.txt); \
     else \

app.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 
 from pathlib import Path
+from types import SimpleNamespace
 import argparse
 import logging
 import json
@@ -39,37 +40,37 @@ def main():
 
     # load config file
     with open(args.config) as cfg_file:
-        config = json.load(cfg_file, cls=ConfigDecoder)
+        config = json.load(cfg_file, cls=ConfigDecoder, object_hook=lambda d: SimpleNamespace(**d))
+
+    stream = fastmot.VideoIO(config.resize_to, args.input_uri, args.output_uri, **vars(config.stream_cfg))
 
     mot = None
     log = None
-    stream = fastmot.VideoIO(config['resize_to'], config['video_io'], args.input_uri, args.output_uri)
-
     if args.mot:
         draw = args.gui or args.output_uri is not None
-        mot = fastmot.MOT(config['resize_to'], config['mot'], draw=draw, verbose=args.verbose)
+        mot = fastmot.MOT(config.resize_to, **vars(config.mot_cfg), draw=draw)
         mot.reset(stream.cap_dt)
         if args.log is not None:
             Path(args.log).parent.mkdir(parents=True, exist_ok=True)
             log = open(args.log, 'w')
     if args.gui:
-        cv2.namedWindow("Video", cv2.WINDOW_AUTOSIZE)
+        cv2.namedWindow('Video', cv2.WINDOW_AUTOSIZE)
 
     logger.info('Starting video capture...')
     stream.start_capture()
     try:
         with Profiler('app') as prof:
-            while not args.gui or cv2.getWindowProperty("Video", 0) >= 0:
+            while not args.gui or cv2.getWindowProperty('Video', 0) >= 0:
                 frame = stream.read()
                 if frame is None:
                     break
 
                 if args.mot:
                     mot.step(frame)
                     if log is not None:
-                        for track in mot.visible_tracks:
-                            tl = track.tlbr[:2] / config['resize_to'] * stream.resolution
-                            br = track.tlbr[2:] / config['resize_to'] * stream.resolution
+                        for track in mot.visible_tracks():
+                            tl = track.tlbr[:2] / config.resize_to * stream.resolution
+                            br = track.tlbr[2:] / config.resize_to * stream.resolution
                             w, h = br - tl + 1
                             log.write(f'{mot.frame_count},{track.trk_id},{tl[0]:.6f},{tl[1]:.6f},'
                                       f'{w:.6f},{h:.6f},-1,-1,-1\n')

cfg/mot.json

@@ -1,78 +1,80 @@
 {   
     "resize_to": [1280, 720],
 
-    "video_io": {
+    "stream_cfg": {
         "resolution": [1920, 1080],
         "frame_rate": 30,
         "buffer_size": 10
     },
 
-    "mot": {
+    "mot_cfg": {
         "detector_type": "YOLO",
         "detector_frame_skip": 5,
 
-        "ssd_detector": {
+        "ssd_detector_cfg": {
             "model": "SSDInceptionV2",
             "class_ids": [1],
             "tile_overlap": 0.25,
             "tiling_grid": [4, 2],
             "conf_thresh": 0.5,
-            "max_area": 130000,
-            "merge_thresh": 0.6
+            "merge_thresh": 0.6,
+            "max_area": 120000
         },
-        "yolo_detector": {
+        "yolo_detector_cfg": {
             "model": "YOLOv4",
             "class_ids": [1],
             "conf_thresh": 0.25,
+            "nms_thresh": 0.5,
             "max_area": 800000,
-            "nms_thresh": 0.5
+            "min_aspect_ratio": 1.2
         },
-        "public_detector": {
-            "sequence": "eval/data/MOT20-03",
+        "public_detector_cfg": {
+            "sequence_path": "MOT20/train/MOT20-01",
             "conf_thresh": 0.5,
             "max_area": 800000
         },
-        "feature_extractor": {
+        "feature_extractor_cfg": {
             "model": "OSNet025",
             "batch_size": 16
         },
 
-        "multi_tracker": {
+        "tracker_cfg": {
             "max_age": 6,
             "age_penalty": 2,
-            "age_weight": 0.1,
-            "motion_weight": 0.02,
-            "max_feat_cost": 0.9,
+            "motion_weight": 0.2,
+            "max_assoc_cost": 0.8,
             "max_reid_cost": 0.6,
             "iou_thresh": 0.4,
-            "duplicate_iou": 0.8,
+            "duplicate_thresh": 0.8,
+            "occlusion_thresh": 0.7,
             "conf_thresh": 0.5,
-            "lost_buf_size": 50,
+            "confirm_hits": 1,
+            "history_size": 50,
 
-            "kalman_filter": {
+            "kalman_filter_cfg": {
                 "std_factor_acc": 2.25,
                 "std_offset_acc": 78.5,
                 "std_factor_det": [0.08, 0.08],
-                "std_factor_flow": [0.14, 0.14],
+                "std_factor_klt": [0.14, 0.14],
                 "min_std_det": [4.0, 4.0],
-                "min_std_flow": [5.0, 5.0],
+                "min_std_klt": [5.0, 5.0],
                 "init_pos_weight": 5,
-                "init_vel_weight": 15,
+                "init_vel_weight": 12,
                 "vel_coupling": 0.6,
                 "vel_half_life": 2
             },
 
-            "flow": {
+            "flow_cfg": {
                 "bg_feat_scale_factor": [0.1, 0.1],
                 "opt_flow_scale_factor": [0.5, 0.5],
-                "feature_density": 0.005,
+                "feat_density": 0.005,
                 "feat_dist_factor": 0.06,
                 "ransac_max_iter": 500,
                 "ransac_conf": 0.99,
                 "max_error": 100,
                 "inlier_thresh": 4,
                 "bg_feat_thresh": 10,
-                "target_feat_params": {
+                "obj_feat_params": {
                     "maxCorners": 1000,
                     "qualityLevel": 0.06,
                     "blockSize": 3
@@ -83,6 +85,15 @@
                     "criteria": [3, 10, 0.03]
                 }
             }
+        },
+
+        "visualizer_cfg": {
+            "draw_detections": false,
+            "draw_confidence": false,
+            "draw_covariance": false,
+            "draw_klt": false,
+            "draw_obj_flow": false,
+            "draw_bg_flow": false
         }
     }
 }