Focustrack Eval

focustrack-eval

FocusTrack: A Self-Adaptive Local Sampling Algorithm for Efficient Anti-UAV Tracking — Wang et al. (2025) (arXiv:2504.13604, 2025)

What this evaluates

Evaluates visual object tracking performance specifically for anti-UAV scenarios, probing a model's ability to maintain target localization under abrupt camera motion, extreme scale variations, and small target sizes in thermal infrared imagery.

Datasets

• AntiUAV — total 318; splits: test (318)
• AntiUAV410 — total 410; splits: train (200), val (90), test (120)

Metrics

• AUC (primary) — range: percent
  • Area under the curve of the success plot, computed by averaging the Intersection over Union (IoU) scores between predicted and ground truth bounding boxes across a dense range of thresholds.
• P — range: percent
  • Precision score at 20 pixels, calculated as the percentage of frames where the Euclidean distance between the predicted and ground truth center coordinates is ≤ 20 pixels.
• PNorm — range: percent
  • Normalized precision, which divides the center location error by the target's dimensions to account for scale variations, then computes the percentage of frames where the normalized error is ≤ 0.5.
• SA — range: percent
  • State accuracy, which jointly evaluates bounding box localization accuracy and target visibility prediction to handle occlusion and out-of-view scenarios.

Input / output format

Input: Paired thermal infrared video frames: a 128×128 template frame and a 256×256 search frame, along with ground-truth bounding box annotations during training.

Output: Predicted bounding box coordinates (center-based), CLS token logits for search region adjustment, and optional segmentation mask from the ATM module.

Scoring recipe

def compute_metrics(pred_boxes, gt_boxes, vis_flags=None):
    ious = [iou(p, g) for p, g in zip(pred_boxes, gt_boxes)]
    auc = np.mean(ious)  # AUC approximated by mean IoU over thresholds
    center_errors = [euclidean_dist(p.center, g.center) for p, g in zip(pred_boxes, gt_boxes)]
    p = 100 * sum(1 for e in center_errors if e <= 20) / len(center_errors)
    pnorm = 100 * sum(1 for e, g in zip(center_errors, gt_boxes) if e / g.size <= 0.5) / len(center_errors)
    sa = compute_state_accuracy(pred_boxes, gt_boxes, vis_flags)  # combines loc & vis
    return {'AUC': auc, 'P': p, 'PNorm': pnorm, 'SA': sa}

Common pitfalls

• Evaluating on the visible modality instead of the specified thermal infrared modality for AntiUAV.
• Comparing against baseline trackers that were not retrained on the AntiUAV410 training set, which violates the paper's fair comparison protocol.
• Failing to disable the Hanning window penalty when the search region is dynamically expanded or shrunk during inference.

Evidence (verbatim from paper)

Tracking performance is evaluated using four key metrics: success rate, precision, normalized precision, and state accuracy[[5]]. The success rate uses the IoU between predicted and ground truth boxes, with the Area Under Curve (AUC) of the success plot serving as the overall accuracy metric. The precision measures center location error, with precision score (P) at 20 pixels as the location accuracy measure. Normalized precision (Pnorm) accounts for target size variations by normalizing the center error with target dimensions. Additionally, state accuracy (SA), proposed in[[5]], evaluates both localization accuracy and visibility prediction, providing a more comprehensive assessment of tracking performance in Anti-UAV tracking scenarios.

Citation

@misc{wang2025focustrack,
  title={FocusTrack: A Self-Adaptive Local Sampling Algorithm for Efficient Anti-UAV Tracking},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2504.13604}
}

• arXiv: 2504.13604