Merge pull request #145 from tryolabs/default-optimized-filter

Set optimized filter as default
tryolabs · Aug 16, 2022 · 4b5081a · 4b5081a
2 parents 3bdaf04 + 3d373e2
commit 4b5081a
Show file tree

Hide file tree

Showing 6 changed files with 35 additions and 31 deletions.
diff --git a/demos/motmetrics4norfair/motmetrics4norfair.py b/demos/motmetrics4norfair/motmetrics4norfair.py
@@ -4,6 +4,7 @@
 import numpy as np
 
 from norfair import Tracker, drawing, metrics, video
+from norfair.filter import FilterPyKalmanFilterFactory
 
 frame_skip_period = 1
 detection_threshold = 0.01
@@ -124,6 +125,7 @@ def keypoints_distance(detected_pose, tracked_pose):
         detection_threshold=detection_threshold,
         pointwise_hit_counter_max=pointwise_hit_counter_max,
         hit_counter_max=hit_counter_max,
+        filter_factory=FilterPyKalmanFilterFactory(),
     )
 
     # Initialize accumulator for this video

diff --git a/demos/motmetrics4norfair/motmetrics4norfair_xyah.py b/demos/motmetrics4norfair/motmetrics4norfair_xyah.py
@@ -7,6 +7,7 @@
 sys.path.append('../../norfair')
 from tracker import Tracker
 from norfair import drawing, metrics, video, Detection
+from norfair.filter import FilterPyKalmanFilterFactory
 
 frame_skip_period = 1
 detection_threshold = 0.01
@@ -129,6 +130,7 @@ def iou_xyah(detected_pose, tracked_pose):
         detection_threshold=detection_threshold,
         pointwise_hit_counter_max=pointwise_hit_counter_max,
         hit_counter_max=hit_counter_max,
+        filter_factory=FilterPyKalmanFilterFactory(),
     )
 
     # Initialize accumulator for this video

diff --git a/docs/README.md b/docs/README.md
@@ -14,7 +14,7 @@ The class in charge of performing the tracking of the detections produced by the
 - `initialization_delay (optional)`: The argument `initialization_delay` determines by how large the object's hit counter must be in order to be considered as initialized, and get returned to the user as a real object. It must be smaller than `hit_counter_max` or otherwise the object would never be initialized. If set to 0, objects will get returned to the user as soon as they are detected for the first time, which can be problematic as this can result in objects appearing and immediately dissapearing. Defaults to `hit_counter_max / 2`.
 - `pointwise_hit_counter_max (optional)`: Each tracked object keeps track of how often the points it's tracking have been getting matched. Points that are getting matched (`pointwise_hit_counter > 0`) are said to be live, and points which aren't (`pointwise_hit_counter = 0`) are said to not be live. This is used to determine things like which individual points in a tracked object get drawn by [`draw_tracked_objects`](#draw_tracked_objects) and which don't. This argument (`pointwise_hit_counter_max`) defines how large the inertia for each point of a tracker can grow. Defaults to `5`.
 - `detection_threshold (optional)`: Sets the threshold at which the scores of the points in a detection being fed into the tracker must dip below to be ignored by the tracker. Defaults to `0`.
-- `filter_factory (optional)`: This parameter can be used to change what filter the [`TrackedObject`](#trackedobject) instances created by the tracker will use. Defaults to [`FilterPyKalmanFilterFactory()`](#filterpykalmanfilterfactory).
+- `filter_factory (optional)`: This parameter can be used to change what filter the [`TrackedObject`](#trackedobject) instances created by the tracker will use. Defaults to [`OptimizedKalmanFilterFactory()`](#optimizedkalmanfilterfactory).
 - `past_detections_length`: How many past detections to save for each tracked object. Norfair tries to distribute these past detections uniformly through the object's lifetime so they're more representative. Very useful if you want to add metric learning to your model, as you can associate an embedding to each detection and access them in your distance function. Defaults to `4`.
 
 ### Tracker.update

diff --git a/norfair/tracker.py b/norfair/tracker.py
@@ -4,9 +4,8 @@
 import numpy as np
 from rich import print
 
-from norfair.distances import get_distance_by_name
-
-from .filter import FilterPyKalmanFilterFactory
+from .distances import get_distance_by_name
+from .filter import OptimizedKalmanFilterFactory
 from .utils import validate_points
 
 
@@ -19,7 +18,7 @@ def __init__(
         initialization_delay: Optional[int] = None,
         pointwise_hit_counter_max: int = 4,
         detection_threshold: float = 0,
-        filter_factory: "FilterPyKalmanFilterFactory" = FilterPyKalmanFilterFactory(),
+        filter_factory: "OptimizedKalmanFilterFactory" = OptimizedKalmanFilterFactory(),
         past_detections_length: int = 4,
     ):
         self.tracked_objects: Sequence["TrackedObject"] = []

diff --git a/tests/metrics.txt b/tests/metrics.txt
@@ -1,23 +1,23 @@
-                IDF1   IDP   IDR  Rcll  Prcn   GT  MT  PT  ML    FP     FN  IDs    FM  MOTA  MOTP  IDt  IDa IDm
-MOT17-04-FRCNN 55.7% 73.9% 44.6% 56.3% 93.2%   83  18  43  22  1962  20778   90   104 52.0% 0.107    5   88   3
-MOT17-11-FRCNN 54.6% 68.6% 45.3% 61.5% 93.1%   75  19  33  23   431   3631   64    61 56.3% 0.101   15   52   5
-MOT17-05-SDP   58.5% 67.6% 51.5% 66.7% 87.6%  133  32  81  20   653   2301  134   140 55.4% 0.165   48   99  15
-MOT17-10-FRCNN 39.9% 44.7% 36.0% 61.0% 75.8%   57  16  35   6  2504   5013  319   313 39.0% 0.173  125  193  10
-MOT17-05-DPM   36.1% 57.1% 26.4% 38.0% 82.2%  133  10  58  65   570   4291   96    96 28.3% 0.242   33   72  12
-MOT17-11-DPM   38.8% 49.7% 31.8% 54.2% 84.5%   75  12  24  39   935   4321   88    64 43.4% 0.217   20   69   4
-MOT17-04-DPM   29.0% 43.1% 21.9% 42.5% 83.6%   83   7  44  32  3965  27336  401   420 33.3% 0.210   43  352   1
-MOT17-10-SDP   47.4% 51.8% 43.6% 74.2% 88.1%   57  30  24   3  1292   3316  308   289 61.7% 0.198  139  165   6
-MOT17-11-SDP   55.4% 61.0% 50.8% 75.8% 91.1%   75  34  30  11   697   2285  103   101 67.3% 0.140   43   69  13
-MOT17-13-SDP   54.4% 66.4% 46.1% 57.0% 82.1%  110  45  28  37  1444   5008  229   209 42.6% 0.201   73  166  18
-MOT17-09-DPM   33.4% 37.7% 30.0% 59.9% 75.4%   26   4  18   4  1042   2137  119   113 38.1% 0.262   41   81   3
-MOT17-13-DPM   21.3% 59.9% 12.9% 18.0% 83.5%  110   5  28  77   416   9543  120   125 13.4% 0.268   35   94   9
-MOT17-05-FRCNN 54.7% 72.0% 44.1% 54.7% 89.2%  133  24  68  41   457   3136   95    96 46.7% 0.181   44   68  17
-MOT17-02-FRCNN 33.9% 53.8% 24.7% 36.6% 79.7%   62   7  26  29  1736  11783  119   131 26.6% 0.134   31   95   8
-MOT17-09-SDP   53.2% 63.8% 45.6% 67.6% 94.6%   26  12  14   0   204   1726   52    55 62.8% 0.130   21   37   6
-MOT17-09-FRCNN 54.0% 72.4% 43.1% 58.6% 98.5%   26   7  17   2    49   2207   40    39 56.9% 0.095   14   30   4
-MOT17-02-SDP   34.4% 42.9% 28.7% 51.0% 76.1%   62  11  39  12  2979   9103  268   284 33.5% 0.182   82  190   8
-MOT17-02-DPM   16.2% 41.0% 10.1% 20.2% 81.6%   62   5  14  43   843  14834  111   112 15.0% 0.246   19   91   1
-MOT17-13-FRCNN 47.0% 52.9% 42.3% 58.8% 73.6%  110  29  57  24  2455   4802  371   366 34.5% 0.185  117  264  20
-MOT17-10-DPM   28.0% 45.8% 20.2% 37.3% 84.6%   57   6  20  31   871   8051  127   154 29.5% 0.248   15  118   6
-MOT17-04-SDP   66.2% 74.7% 59.5% 77.6% 97.4%   83  48  26   9  1001  10672  225   254 75.0% 0.132   89  136   5
-OVERALL        45.3% 59.5% 36.6% 53.6% 87.2% 1638 381 727 530 26506 156274 3479  3526 44.7% 0.163 1052 2529 174
+                IDF1   IDP   IDR  Rcll  Prcn   GT  MT  PT  ML    FP     FN  IDs    FM  MOTA  MOTP IDt  IDa IDm
+MOT17-10-DPM   27.4% 44.8% 19.8% 37.5% 84.9%   57   6  19  32   856   8024  127   153 29.8% 0.249  19  116   9
+MOT17-04-SDP   66.2% 74.6% 59.4% 77.5% 97.4%   83  49  25   9  1003  10680  232   257 74.9% 0.132  94  136   4
+MOT17-11-SDP   55.2% 60.8% 50.6% 75.7% 90.9%   75  34  30  11   719   2297  112   105 66.9% 0.141  47   73  13
+MOT17-13-FRCNN 48.1% 54.3% 43.1% 59.3% 74.7%  110  33  53  24  2334   4735  342   331 36.3% 0.184 115  232  17
+MOT17-05-FRCNN 55.2% 72.9% 44.4% 54.9% 90.0%  133  23  69  41   420   3123   78    84 47.7% 0.180  39   55  16
+MOT17-02-SDP   34.4% 42.8% 28.7% 50.9% 75.9%   62  11  38  13  3003   9122  272   290 33.3% 0.184  72  203   7
+MOT17-13-SDP   56.0% 68.5% 47.4% 57.6% 83.3%  110  46  26  38  1349   4934  164   163 44.6% 0.200  47  126  21
+MOT17-10-SDP   47.5% 51.9% 43.7% 74.1% 88.0%   57  30  24   3  1295   3323  283   286 61.8% 0.198 119  157   5
+MOT17-02-FRCNN 33.9% 53.8% 24.7% 36.5% 79.5%   62   7  26  29  1751  11796  124   136 26.4% 0.134  37   95   9
+MOT17-13-DPM   22.1% 62.3% 13.4% 18.5% 85.8%  110   6  31  73   355   9490   90   100 14.7% 0.267  14   81   5
+MOT17-04-FRCNN 55.7% 73.9% 44.7% 56.3% 93.2%   83  18  43  22  1959  20775   90   104 52.0% 0.107   5   88   3
+MOT17-04-DPM   29.2% 43.3% 22.0% 42.5% 83.5%   83   7  44  32  3985  27365  406   432 33.2% 0.211  48  352   1
+MOT17-09-FRCNN 54.0% 72.4% 43.1% 58.5% 98.5%   26   7  17   2    49   2209   40    39 56.8% 0.096  15   29   4
+MOT17-11-DPM   39.2% 50.3% 32.2% 54.2% 84.7%   75  12  24  39   927   4321   87    63 43.5% 0.217  20   68   4
+MOT17-09-DPM   33.4% 37.7% 30.0% 59.7% 75.1%   26   4  17   5  1052   2147  120   117 37.7% 0.263  41   82   3
+MOT17-05-SDP   58.1% 67.3% 51.1% 66.8% 87.9%  133  32  81  20   637   2299  134   133 55.6% 0.165  51   98  18
+MOT17-09-SDP   53.2% 63.8% 45.6% 67.7% 94.6%   26  12  14   0   204   1722   54    56 62.8% 0.130  22   38   6
+MOT17-11-FRCNN 54.3% 68.3% 45.1% 61.4% 92.9%   75  18  34  23   443   3639   65    62 56.1% 0.102  16   54   6
+MOT17-10-FRCNN 39.8% 44.7% 35.9% 60.9% 75.7%   57  14  37   6  2507   5020  310   306 39.0% 0.172 122  186   9
+MOT17-05-DPM   36.5% 57.9% 26.6% 38.1% 82.9%  133  11  58  64   544   4281   90   100 28.9% 0.243  33   68  13
+MOT17-02-DPM   16.2% 40.9% 10.1% 20.2% 81.4%   62   5  14  43   856  14833  113   109 15.0% 0.246  20   92   1
+OVERALL        45.4% 59.7% 36.7% 53.7% 87.3% 1638 385 724 529 26248 156135 3333  3426 44.9% 0.164 996 2429 174
diff --git a/tests/test_mot_metrics.py b/tests/test_mot_metrics.py
@@ -3,7 +3,7 @@
 import numpy as np
 import pandas as pd
 
-from norfair import Tracker, metrics
+from norfair import Tracker, metrics, FilterPyKalmanFilterFactory
 
 DATASET_PATH = "train"
 MOTA_ERROR_THRESHOLD = 0.0
@@ -80,6 +80,7 @@ def test_mot_metrics():
             detection_threshold=DETECTION_THRESHOLD,
             pointwise_hit_counter_max=POINTWISE_HIT_COUNTER_MAX,
             hit_counter_max=HIT_COUNTER_MAX,
+            filter_factory=FilterPyKalmanFilterFactory(),
         )
 
         # Initialize accumulator for this video
@@ -103,8 +104,8 @@ def test_mot_metrics():
     # Unify the scores to be able to compare them. new metrics is the percentage
     # expressed between 0 and 1, the previous metrics have the percentage as a string
     # with the % character at the end
-    new_overall_mota = new_metrics.loc["OVERALL", "mota"] * 100
-    previous_overall_mota = float(previous_metrics.loc["OVERALL", "mota"][:-1])
+    new_overall_mota = np.around(new_metrics.loc["OVERALL", "mota"] * 100, 1)
+    previous_overall_mota = np.around(float(previous_metrics.loc["OVERALL", "mota"][:-1]), 1)
 
     accumulator.print_metrics()
     assert new_overall_mota >= previous_overall_mota * (1 - MOTA_ERROR_THRESHOLD), f"New overall MOTA score: {new_overall_mota} is too low, previous overall MOTA score: {previous_overall_mota}"