[Python] VOT + YOLO&BoT-SORT: Evaluating Multi-Object Tracker on Single-Object Benchmark -...

I'm trying to evaluate a multi-object tracker on a single-object tracking benchmark on Python. My approach involves selecting the target ID from the bounding box in the first frame that best matches the initial bounding box in the first frame. I'm using a get_best_match() function for this. Since YOLO might struggle with small or unlearned objects, I'm specifically testing on video sequences where YOLO is expected to detect the target.

However, my code stops after 2 frames and I'm having trouble pinpointing the root cause.

The code:

import numpy as np

import vot
import sys
import cv2

from ultralytics import YOLO

def iou(box_a, box_b):
# Compute intersection-over-union for two boxes [x1, y1, x2, y2]
x_a = max(box_a[0], box_b[0])
y_a = max(box_a[1], box_b[1])
x_b = min(box_a[2], box_b[2])
y_b = min(box_a[3], box_b[3])
interArea = max(0, x_b - x_a) * max(0, y_b - y_a)
boxAArea = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
boxBArea = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
IoU = interArea / float(boxAArea + boxBArea - interArea)
return IoU

def get_best_match(selection_box, boxes):
# selection_box: [x1, y1, x2, y2]
# boxes: iterable of box objects with .xyxy attribute
max_iou = 0
best_id = None
best_box = None
for box in boxes:
box_xyxy = box.xyxy.cpu().numpy().tolist()[0]
iou_score = iou(selection_box, box_xyxy)
if iou_score > max_iou:
max_iou = iou_score
best_id = int(box.id.item()) if box.id is not None else None
best_box = box
if max_iou == 0 and len(boxes) > 0:
# Fallback: pick closest box by center
sel_cx = (selection_box[0] + selection_box[2]) / 2
sel_cy = (selection_box[1] + selection_box[3]) / 2
min_dist = float('inf')
for box in boxes:
box_xyxy = box.xyxy.cpu().numpy().tolist()[0]
bx = (box_xyxy[0] + box_xyxy[2]) / 2
by = (box_xyxy[1] + box_xyxy[3]) / 2
dist = (sel_cx - bx) ** 2 + (sel_cy - by) ** 2
if dist < min_dist:
min_dist = dist
best_id = int(box.id.item()) if box.id is not None else None
best_box = box
return best_id, best_box



detector = YOLO('/home/uuu/jupyter_projects/MyProjects/baka/yolo11n.pt')

handle = vot.VOT("rectangle")
selection = handle.region()
selection_box = [selection.x, selection.y, selection.x + selection.width, selection.y + selection.height]

max_iou = 0
target_id = None
score = 0
x1, y1, x2, y2 = 0, 0, 0, 0

results = detector.track(handle.frame(), tracker='botsort.yaml', stream=True, persist=True)

for result in results:
boxes = result.boxes
try:
if target_id is None:
for box in boxes:
target_id, _ = get_best_match(selection_box, boxes)
if target_id is None:
print("Target ID not found -_-")
handle.report(vot.Rectangle(0, 0, 0, 0), 0.0)

found = False
for box in boxes:
curr_id = int(box.id.item()) if box.id is not None else None
if curr_id == target_id:
box_xyxy = box.xyxy.cpu().numpy().tolist()[0]
x1, y1, x2, y2 = map(int, box_xyxy)
score = float(box.conf.item())
found = True
break

if found:
handle.report(vot.Rectangle(x1, y1, x2 - x1, y2 - y1), score)
else:
handle.report(vot.Rectangle(0, 0, 0, 0), 0.0)

except Exception as e:
# Always respond, even if something goes wrong
handle.report(vot.Rectangle(0, 0, 0, 0), 0.0)
print(f"Error: {e}", file=sys.stderr)


The output: output

Any ideas on how I can figure out why my code stops after the second frame? Or if you see any obvious mistakes in my code, please let me know.

I tried to detect objects in each frame individually and then provide these detections to the tracker, aiming to use handle.frame() on each frame as in the VOT example. The issue I faced was understanding how to properly feed those detections into the tracker.

Continue reading...