Feat/core/oc sort alex by AlexBodner · Pull Request #162 · roboflow/trackers

AlexBodner · 2026-01-20T17:47:20Z

What does this PR do?

OC-SORT in process

Type of Change

New feature (non-breaking change that adds functionality)

Testing

Benchmark soccernet: https://colab.research.google.com/drive/1zzBf4xBSN9KBSDUTDQnMy7Zvny2DOeJL?usp=sharing

…stead of track kf state

…ks before

…trackers into feat/core/oc-sort-alex

tstanczyk95 · 2026-01-28T17:49:18Z

docs/trackers/core/ocsort/tracker.md

+# OC-SORT
+
+[![arXiv](https://img.shields.io/badge/arXiv-2203.14360-b31b1b.svg)](https://arxiv.org/abs/2203.14360)
+[![colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-track-objects-with-sort-tracker.ipynb)


Remember to update it later for OC-SORT

If you look at the current docs, for example the ByteTrack page, you will see badges removed from the top for now. https://trackers.roboflow.com/develop/trackers/bytetrack/

Instead, a Tutorials section appears on the main page, with links to Google Colab notebooks. https://trackers.roboflow.com/develop/

tstanczyk95 · 2026-01-28T17:50:40Z

docs/trackers/core/ocsort/tracker.md

+    1. Observation-Centre Re-Update (ORU): runs a predict-update loop with a 'virtual trajectory'
+        depending on the last observation and new observation when a track is re-activated after being lost.
+    2. Observation-Centric Momentum (OCM): incorporate the direction consistency of tracks in the cost matrix for the association.
+    3. Observation-centric Recovery (OCR): a second-stage association step between the last observation of unmatched tracks and the unmatched observations after the usual association. It attempts to recover tracks that were lost due to object stopping or short-term occlusion. Uses only IoU.


"Uses only IoIU" as a separate line?

I think that is a problem of the visualization here maybe, because I see it in the same line

tstanczyk95 · 2026-01-28T17:51:30Z

docs/trackers/core/ocsort/tracker.md

+
+## Overview
+
+OC-SORT remains Simple, Online, and Real-Time ([SORT](../sort/tracker.md)) but improves robustness during occlusion and non-linear motion.


"remains... like SORT". Otherwise, it sounds confusing

tstanczyk95 · 2026-01-28T17:53:37Z

trackers/core/ocsort/tracker.py

+            improving tracking through occlusions, but may increase the possibility
+            of ID switching for objects with similar appearance.
+        frame_rate (float): Frame rate of the video (frames per second).
+            Used to calculate the maximum time a track can be lost.


specify that we talk about the number of frames: for how many frames a track/tracked entity can be lost

I didnt understand the comment, can you explain further?

tstanczyk95 · 2026-01-28T18:12:35Z

trackers/core/ocsort/tracker.py

+                form [x1, y1, x2, y2].
+        """
+        for detection_idx in unmatched_detections:
+            new_tracker = OCSORTTracklet(detections.xyxy[detection_idx])


This might be super confusing in the future. I would opt for creating a new track or new tracklet, not a new tracker.

Looks like new traklet to me.

yes, agree. this came from me basing the code on SORT

also, one comment on this, sv.Detections have the attribute 'tracker_id', so we have this colliding term there as well.

tstanczyk95 · 2026-01-28T18:39:02Z

trackers/utils/ocsort_utils.py

+    for t, tracklet in enumerate(tracklets):
+        if tracklet.previous_to_last_observation is None:  # if there is no previous box
+            continue
+        # In case the track is lost we use the last known box


Is it by the original OC-SORT design? It might introduce some measurement noise and lead towards a drift.

yes, if there is not enough information to calculate a direction then that component is 0

tstanczyk95

The implementation code looks overall good with a few remarks from my side. In any case, we will need to discuss it online to clarify some parts and design choices as well as ensure about the parameters.

SkalskiP · 2026-01-28T22:29:07Z

docs/trackers/core/ocsort/tracker.md

+# OC-SORT
+
+[![arXiv](https://img.shields.io/badge/arXiv-2203.14360-b31b1b.svg)](https://arxiv.org/abs/2203.14360)
+[![colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-track-objects-with-sort-tracker.ipynb)


If you look at the current docs, for example the ByteTrack page, you will see badges removed from the top for now. https://trackers.roboflow.com/develop/trackers/bytetrack/

Instead, a Tutorials section appears on the main page, with links to Google Colab notebooks. https://trackers.roboflow.com/develop/

SkalskiP · 2026-01-28T22:32:35Z

docs/trackers/core/ocsort/tracker.md

+OC-SORT remains Simple, Online, and Real-Time ([SORT](../sort/tracker.md)) but improves robustness during occlusion and non-linear motion.
+It recognizes limitations from SORT and the linear motion assumption of the Kalman filter, and adds three
+mechanisms to enhance tracking:
+    1. Observation-Centre Re-Update (ORU): runs a predict-update loop with a 'virtual trajectory'
+        depending on the last observation and new observation when a track is re-activated after being lost.
+    2. Observation-Centric Momentum (OCM): incorporate the direction consistency of tracks in the cost matrix for the association.
+    3. Observation-centric Recovery (OCR): a second-stage association step between the last observation of unmatched tracks and the unmatched observations after the usual association. It attempts to recover tracks that were lost due to object stopping or short-term occlusion. Uses only IoU.


Let’s keep the text simple, like in the ByteTrack docs. https://trackers.roboflow.com/develop/trackers/bytetrack/ No lists. A single paragraph of text.

SkalskiP · 2026-01-28T22:33:45Z

docs/trackers/core/ocsort/tracker.md

+        depending on the last observation and new observation when a track is re-activated after being lost.
+    2. Observation-Centric Momentum (OCM): incorporate the direction consistency of tracks in the cost matrix for the association.
+    3. Observation-centric Recovery (OCR): a second-stage association step between the last observation of unmatched tracks and the unmatched observations after the usual association. It attempts to recover tracks that were lost due to object stopping or short-term occlusion. Uses only IoU.
+


We should have Benchmarks section here.

SkalskiP · 2026-01-28T22:37:31Z

docs/trackers/core/ocsort/tracker.md

+## Examples
+
+=== "inference"
+
+    ```python hl_lines="2 5 12"
+    import supervision as sv
+    from trackers import OCSORTTracker
+    from inference import get_model
+
+    tracker = OCSORTTracker()
+    model = get_model(model_id="rfdetr-medium")
+    annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER)
+
+    def callback(frame, _):
+        result = model.infer(frame)[0]
+        detections = sv.Detections.from_inference(result)
+        detections = tracker.update(detections)
+        return annotator.annotate(frame, detections, labels=detections.tracker_id)
+
+    sv.process_video(
+        source_path="<INPUT_VIDEO_PATH>",
+        target_path="<OUTPUT_VIDEO_PATH>",
+        callback=callback,
+    )
+    ```
+
+=== "rf-detr"
+
+    ```python hl_lines="2 5 11"
+    import supervision as sv
+    from trackers import OCSORTTracker
+    from rfdetr import RFDETRMedium
+
+    tracker = OCSORTTracker()
+    model = RFDETRMedium()
+    annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER)
+
+    def callback(frame, _):
+        detections = model.predict(frame)
+        detections = tracker.update(detections)
+        return annotator.annotate(frame, detections, labels=detections.tracker_id)
+
+    sv.process_video(
+        source_path="<INPUT_VIDEO_PATH>",
+        target_path="<OUTPUT_VIDEO_PATH>",
+        callback=callback,
+    )
+    ```
+
+=== "ultralytics"
+
+    ```python hl_lines="2 5 12"
+    import supervision as sv
+    from trackers import OCSORTTracker
+    from ultralytics import YOLO
+
+    tracker = OCSORTTracker()
+    model = YOLO("yolo11m.pt")
+    annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER)
+
+    def callback(frame, _):
+        result = model(frame)[0]
+        detections = sv.Detections.from_ultralytics(result)
+        detections = tracker.update(detections)
+        return annotator.annotate(frame, detections, labels=detections.tracker_id)
+
+    sv.process_video(
+        source_path="<INPUT_VIDEO_PATH>",
+        target_path="<OUTPUT_VIDEO_PATH>",
+        callback=callback,
+    )
+    ```
+
+=== "transformers"
+
+    ```python hl_lines="3 6 28"
+    import torch
+    import supervision as sv
+    from trackers import OCSORTTracker
+    from transformers import RTDetrV2ForObjectDetection, RTDetrImageProcessor
+
+    tracker = OCSORTTracker()
+    processor = RTDetrImageProcessor.from_pretrained("PekingU/rtdetr_v2_r18vd")
+    model = RTDetrV2ForObjectDetection.from_pretrained("PekingU/rtdetr_v2_r18vd")
+    annotator = sv.LabelAnnotator(text_position=sv.Position.CENTER)
+
+    def callback(frame, _):
+        inputs = processor(images=frame, return_tensors="pt")
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        h, w, _ = frame.shape
+        results = processor.post_process_object_detection(
+            outputs,
+            target_sizes=torch.tensor([(h, w)]),
+            threshold=0.5
+        )[0]
+
+        detections = sv.Detections.from_transformers(
+            transformers_results=results,
+            id2label=model.config.id2label
+        )
+
+        detections = tracker.update(detections)
+        return annotator.annotate(frame, detections, labels=detections.tracker_id)
+
+    sv.process_video(
+        source_path="<INPUT_VIDEO_PATH>",
+        target_path="<OUTPUT_VIDEO_PATH>",
+        callback=callback,
+    )
+    ```


Before v2.1.0 release I updated the format of similar examples for SORT and ByteTrack. Please make sure we follow similar pattern here.

SkalskiP · 2026-01-28T22:40:21Z

trackers/core/ocsort/tracker.py

+            to the unmatched observations after the usual association. It attempts to recover tracks that were lost
+            due to object stopping or short-term occlusion.
+    Args:
+        lost_track_buffer (int): Number of frames to buffer when a track is lost.


Let's drop types from docstring, we only keep it in code. We are slowly doing the same in other repos. I did that in trackers before v2.1.0 release. I'm flagging it here, but we would need to do it in the whole PR.

SkalskiP · 2026-01-28T22:55:45Z

trackers/utils/bbox_conversions.py

+    h = bbox[3] - bbox[1]
+    x = bbox[0] + w / 2.0
+    y = bbox[1] + h / 2.0
+    s = w * h  # scale is just area


Comments like this seem redundent, it's a pretty simple math operation. You can point out in docstring. that scale is an area.

SkalskiP · 2026-01-28T22:58:02Z

trackers/utils/bbox_conversions.py

@@ -0,0 +1,34 @@
+import numpy as np


would rename the whole file to converters.py

SkalskiP · 2026-01-28T22:58:48Z

trackers/__init__.py

+from trackers.log import get_logger

-__all__ = ["ByteTrackTracker", "SORTTracker"]
+__all__ = ["OCSORTTracker", "SORTTracker"]


seem like this file is out of date, please make sure all develop chanegs are in this branch

SkalskiP · 2026-01-28T22:58:52Z

mkdocs.yml

-    - ByteTrack: trackers/bytetrack.md
+    - SORT: trackers/core/sort/tracker.md
+    - DeepSORT: trackers/core/deepsort/tracker.md
+    - OC-SORT: trackers/core/ocsort/tracker.md


seem like this file is out of date, please make sure all develop chanegs are in this branch

SkalskiP · 2026-01-28T23:01:03Z

trackers/utils/ocsort_utils.py

@@ -0,0 +1,180 @@
+from copy import deepcopy


why not move this whole file to core/ocsort and rename it to utils.py if it's OC-SORT specific? is this code entierly copy-pasted from their repo?

ok! because before we were putting all tracker specific utils there i continued putting them there, but sounds good to do so

SkalskiP · 2026-01-29T13:00:02Z

@AlexBodner can we close it?

AlexBodner and others added 8 commits January 15, 2026 13:36

initialized ocsort

c881137

OC-SORT first version working

975b6a1

removed prints

f501269

removed unused files

7c692de

added high_conf_det_threshold

327b952

now IoU matrix of OCR step is computed with last track observation in…

cd067fb

…stead of track kf state

small bug fix regarding tracker initialization when ther were no trac…

69c8a8b

…ks before

fix(pre_commit): 🎨 auto format pre-commit hooks

498b09b

AlexBodner requested a review from tstanczyk95 January 22, 2026 17:01

AlexBodner and others added 21 commits January 22, 2026 18:03

fixed pre commit

63e44c4

Merge branch 'feat/core/oc-sort-alex' of https://github.com/roboflow/…

d1aff38

…trackers into feat/core/oc-sort-alex

now ocr doesnt use direction consistency matrix at all

1e1505d

small bug fix that could affect ocr indices

23afd35

small bug fix that could affect ocr indices

ee9e29e

added docs

829d58c

fix(pre_commit): 🎨 auto format pre-commit hooks

43781f0

weighted direction consistency mat by confidence

1577350

weighted direction consistency mat by confidence

6787524

Merge branch 'feat/core/oc-sort-alex' of https://github.com/roboflow/…

bf886df

…trackers into feat/core/oc-sort-alex

Kalman filter state changed, now using xysr

11914be

Fixed circular input

d54e990

fixed mkdocs

bc0c964

changed < by <= to have exact same behaviour"

ba78c66

changed lost to be lsot for 1 frame

86b12b0

rollback

6aa8977

possible fix

74bcd07

fix(pre_commit): 🎨 auto format pre-commit hooks

a02150e

now convert to state coordinates and then interpolate

dd1f6ad

now convert to state coordinates and then interpolate

7ed756e

fix(pre_commit): 🎨 auto format pre-commit hooks

277a259

AlexBodner and others added 8 commits January 27, 2026 11:56

now convert to state coordinates and then interpolate

8a56808

Merge branch 'feat/core/oc-sort-alex' of https://github.com/roboflow/…

206344f

…trackers into feat/core/oc-sort-alex

fix(pre_commit): 🎨 auto format pre-commit hooks

1a71753

now interpolate in xywh and convert to xysq

52b7efd

now interpolate in xywh and convert to xysq

375e8dd

now interpolate in xywh and convert to xysq

43f7328

now interpolate in xywh and convert to xysq

9380e4d

added save kalman filter as function

1d132b0

tstanczyk95 reviewed Jan 28, 2026

View reviewed changes

SkalskiP changed the base branch from main to develop January 28, 2026 22:20

SkalskiP requested changes Jan 28, 2026

View reviewed changes

AlexBodner added 3 commits January 29, 2026 12:25

worked on review comments

39c93c3

changed doc to ocsort

03ca8fe

renamed files

d423795


		## Overview

		OC-SORT remains Simple, Online, and Real-Time ([SORT](../sort/tracker.md)) but improves robustness during occlusion and non-linear motion.

Conversation

AlexBodner commented Jan 20, 2026

What does this PR do?

Type of Change

Testing

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AlexBodner Jan 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tstanczyk95 Jan 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AlexBodner Jan 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tstanczyk95 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SkalskiP commented Jan 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

AlexBodner Jan 29, 2026 •

edited

Loading

tstanczyk95 Jan 28, 2026 •

edited

Loading

AlexBodner Jan 29, 2026 •

edited

Loading