Skip to content

C-BIoU (Cascaded-Buffered IoU)

What is C-BIoU?

C-BIoU builds on the same tracking pipeline as ByteTrack but replaces plain IoU with Buffered IoU (BIoU), expanding boxes before overlap is computed so tracks and detections can still match when motion of the object is fast or boxes barely align. It runs two association passes with a small buffer first and a larger buffer second (buffer_ratio_first and buffer_ratio_second), so only bounding boxes are required. C-BIoU is a strong fit for sports and dance footage where objects move fast and change direction.

How does C-BIoU compare to other trackers?

For comparisons with other trackers, plus default and tuned parameters, see the tracker comparison page.

Dataset HOTA IDF1 MOTA
MOT17 63.0 79.1 77.4
SportsMOT 73.1 72.6 96.7
SoccerNet 82.6 76.6 97.0
DanceTrack 53.8 53.8 90.1

How does C-BIoU work?

C-BIoU keeps the ByteTrack-style association pipeline used in BoT-SORT but replaces plain IoU with Cascaded Buffered IoU at each association step.

First association (b1). High-confidence detections are matched to confirmed and lost tracks using BIoU with buffer_ratio_first (paper b1, small buffer). Costs are fused with detection confidence.

Second association (b2). Remaining tracked tracks (not lost) are matched to low-confidence detections using BIoU with buffer_ratio_second (paper b2, large buffer). In this implementation, this larger buffer corresponds to ByteTrack's recovery stage for unmatched tracks and low-confidence detections.

Unconfirmed association (b1). Leftover high-confidence detections are matched to unconfirmed tracks using the same buffer as pass 1. Unmatched unconfirmed tracks are removed. This step is inherited from ByteTrack lifecycle logic, not the paper's two-buffer cascade.

Track lifecycle. New tracks are initiated and confirmed with a conservative policy (minimum_consecutive_frames) to reduce one-frame false positives. Existing tracks that remain unmatched longer than lost_track_buffer are removed.

Key Parameters

Parameter Purpose Tuning guidance
lost_track_buffer Number of frames to keep an unmatched track alive before deletion. Higher value tolerates longer occlusions but risks false re-association. Use range (10, 30) for most scenes; up to 60 for very long occlusions.
track_activation_threshold Minimum detection confidence required to start a new track. Higher value reduces noisy track creation; lower value retains harder objects. 0.5-0.9 typical depending on detector quality. This does not control low-confidence association, which still discards detections at a fixed 0.1 confidence floor.
minimum_consecutive_frames Number of consecutive matches required before confirming a new track. 1 for immediate activation; 2-3 improves robustness against flicker and false positives.
minimum_iou_threshold_first_assoc Minimum fused BIoU similarity for the first association pass. Lower value helps maintain matches under fast motion; higher value is stricter.
minimum_iou_threshold_second_assoc Minimum BIoU similarity for the second association pass. Usually set to a lower value than the first-pass threshold to recover weak detections without over-matching.
minimum_iou_threshold_unconfirmed_assoc Minimum fused BIoU similarity when associating unconfirmed tracks. Higher value makes tentative tracks harder to confirm spuriously; lower value helps short-lived or noisy objects survive.
high_conf_det_threshold Confidence split between stage-1 and stage-2 detections. 0.5-0.7 common. Higher value shifts more detections to recovery stage; lower value gives stage-1 broader coverage.
buffer_ratio_first Paper b1, small BIoU buffer for the first association pass. Typical range 0.1-0.7. Should be less than buffer_ratio_second.
buffer_ratio_second Paper b2, large BIoU buffer for the second association pass. Typical range 0.2-1.0. Should be greater than buffer_ratio_first.

Buffer ordering (b1 < b2)

Always set buffer_ratio_first < buffer_ratio_second. The cascaded matcher applies the smaller buffer first, then the larger buffer only on pairs that remain unmatched. Reversing the order (b1 ≥ b2) is not consistent with the paper and usually hurts performance.

Frame input is ignored by C-BIoU

CBIoUTracker.update() accepts frame for API consistency with other trackers, but C-BIoU does not use image/frame pixels. If you pass frame with a non-None value, the tracker emits a UserWarning and ignores it.

Run on video, webcam, or RTSP stream

These examples use opencv-python for decoding and display. Replace <SOURCE_VIDEO_PATH>, <WEBCAM_INDEX>, and <RTSP_STREAM_URL> with your inputs. <WEBCAM_INDEX> is usually 0 for the default camera.

import cv2
import supervision as sv
from rfdetr import RFDETRMedium
from trackers import CBIoUTracker

tracker = CBIoUTracker()
model = RFDETRMedium()

box_annotator = sv.BoxAnnotator()
label_annotator = sv.LabelAnnotator()

video_capture = cv2.VideoCapture("<SOURCE_VIDEO_PATH>")
if not video_capture.isOpened():
    raise RuntimeError("Failed to open video source")

while True:
    success, frame_bgr = video_capture.read()
    if not success:
        break

    frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
    detections = model.predict(frame_rgb)
    detections = tracker.update(detections)

    annotated_frame = box_annotator.annotate(frame_bgr, detections)
    annotated_frame = label_annotator.annotate(
        annotated_frame,
        detections,
        labels=detections.tracker_id,
    )

    cv2.imshow("RF-DETR + C-BIoU", annotated_frame)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

video_capture.release()
cv2.destroyAllWindows()

import cv2
import supervision as sv
from rfdetr import RFDETRMedium
from trackers import CBIoUTracker

tracker = CBIoUTracker()
model = RFDETRMedium()

box_annotator = sv.BoxAnnotator()
label_annotator = sv.LabelAnnotator()

video_capture = cv2.VideoCapture("<WEBCAM_INDEX>")
if not video_capture.isOpened():
    raise RuntimeError("Failed to open webcam")

while True:
    success, frame_bgr = video_capture.read()
    if not success:
        break

    frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
    detections = model.predict(frame_rgb)
    detections = tracker.update(detections)

    annotated_frame = box_annotator.annotate(frame_bgr, detections)
    annotated_frame = label_annotator.annotate(
        annotated_frame,
        detections,
        labels=detections.tracker_id,
    )

    cv2.imshow("RF-DETR + C-BIoU", annotated_frame)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

video_capture.release()
cv2.destroyAllWindows()

import cv2
import supervision as sv
from rfdetr import RFDETRMedium
from trackers import CBIoUTracker

tracker = CBIoUTracker()
model = RFDETRMedium()

box_annotator = sv.BoxAnnotator()
label_annotator = sv.LabelAnnotator()

video_capture = cv2.VideoCapture("<RTSP_STREAM_URL>")
if not video_capture.isOpened():
    raise RuntimeError("Failed to open RTSP stream")

while True:
    success, frame_bgr = video_capture.read()
    if not success:
        break

    frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
    detections = model.predict(frame_rgb)
    detections = tracker.update(detections)

    annotated_frame = box_annotator.annotate(frame_bgr, detections)
    annotated_frame = label_annotator.annotate(
        annotated_frame,
        detections,
        labels=detections.tracker_id,
    )

    cv2.imshow("RF-DETR + C-BIoU", annotated_frame)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

video_capture.release()
cv2.destroyAllWindows()

For BIoU mathematics and using BIoU(buffer_ratio=...) on other trackers, see IoU variants. To tune hyperparameters with Optuna, see Hyperparameter tuning.

Reference

Yang, F., Odashima, S., Masui, S., and Jiang, S. (2023). Hard To Track Objects with Irregular Motions and Similar Appearances? Make It Easier by Buffering the Matching Space. WACV 2023. arXiv:2211.14317

Comments