Anchor box ও NMS

NMS threshold — detection-এর critical hyperparameter.

Threshold low (e.g., 0.3):

• Aggressive suppression.
• Touching object-এ একটি miss.
• Output cleaner কিন্তু missed detection.

Threshold high (e.g., 0.7):

• Conservative suppression.
• Duplicate detection survive.
• Output cluttered, but recall high.

Standard 0.5: COCO benchmark — balanced.

Crowded scene strategies:

• Higher threshold (0.7) + post-filter.
• Soft-NMS — gradient decay.
• DETR — set prediction native.
• CrowdHuman dataset — specialized model.

Per-class threshold: people 0.7, car 0.5 — empirical tuning.

Bangladesh use case: traffic — overlap inevitable. Soft-NMS or DETR preferred.

মূল উপলব্ধি: Single-number tuning rarely optimal. Domain-specific NMS strategy পরিগণিত।

Anchor-free — design simplification, not just trick.

Anchor-based complexity:

• Hyperparameter: scale, ratio, count.
• Matching strategy: IoU thresholds.
• Domain-specific tuning.
• Class-agnostic anchor — class-specific scale awkward.

Anchor-free philosophy:

• Each pixel — predict directly: in object? center? size?
• FCOS: per-pixel (center, top, bottom, left, right) distance.
• CenterNet: heatmap of center + size regression.
• No matching IoU needed.

Advantages:

• Fewer hyperparameter.
• Adaptable across dataset.
• Easy multi-task extension (FCOS supports keypoint, segmentation).
• Better small object — direct prediction.

Performance comparison:

• FCOS COCO: 41.5 mAP (vs RetinaNet 39.1).
• YOLOv8 anchor-free: 53.9 mAP (vs YOLOv5 50.7).
• Modern era — anchor-free competitive বা better.

Why anchor-free historically lost:

• Earlier (2018-2019) — anchor-based dominated.
• Required stronger backbone, FPN.
• ২০২০+ — DenseBox, FCOS, CenterNet — proper attention.

মূল উপলব্ধি: Anchor — historical solution, not necessary. Modern shift natural progression. Architecture simplicity often wins.

Soft-NMS-এর math choice — empirical + principled.

Linear decay:

• $s' = s \cdot (1 - \text{IoU})$ if IoU > threshold.
• Simple, no hyperparameter.
• Works OK.

Gaussian decay:

• $s' = s \cdot e^{-\text{IoU}^2 / \sigma}$।
• Smooth, differentiable everywhere.
• Hyperparameter $\sigma$ controls decay rate.
• $\sigma$ small — aggressive. Large — gentle.

Comparison (paper):

• Standard NMS: 39.4 mAP.
• Linear: 40.0.
• Gaussian: 40.8.
• Gaussian wins by 0.8%.

Why Gaussian smoother:

• Continuous decay — no discontinuity at threshold.
• Differentiable — useful in trainable NMS variants.
• Statistically motivated (similar to confidence calibration).

Implementation:

def soft_nms(boxes, scores, sigma=0.5, conf_thresh=0.001):
    indices = np.argsort(-scores)
    keep = []
    while len(indices) > 0:
        i = indices[0]
        keep.append(i)
        ious = iou(boxes[i], boxes[indices[1:]])
        scores[indices[1:]] *= np.exp(-ious**2 / sigma)
        indices = indices[1:][scores[indices[1:]] > conf_thresh]
    return keep

মূল উপলব্ধি: Math choice — both empirical তাত্ত্বিক justification. Soft-NMS — principled refinement of greedy NMS.

DETR — Carion et al. (Facebook AI Research, 2020) — paradigm shift.

Idea:

• Fixed N "object queries" (e.g., 100).
• Each query — predict (class, box) or "no object".
• One-to-one matching with ground-truth.

Hungarian matching:

• Combinatorial optimization — Kuhn-Munkres algorithm.
• Bipartite graph: predictions ↔ ground-truth.
• Edge weight: matching cost (class + bbox + IoU).
• Find perfect matching minimum cost.

Cost function:

$\mathcal{L}_{\text{match}} = \lambda_{\text{cls}} \mathcal{L}_{\text{cls}} + \lambda_{\text{bbox}} \mathcal{L}_{\text{bbox}} + \lambda_{\text{IoU}} \mathcal{L}_{\text{IoU}}$

Effect:

• Each ground-truth → exactly one prediction.
• Other predictions → "no object" (penalty).
• No duplicate by construction.
• NMS unnecessary.

Why end-to-end?

• Loss differentiable through Hungarian matching (with smoothing).
• Training optimizes detection directly.
• No anchor matching heuristic.

DETR drawbacks:

• Slow convergence — 500 epoch typical.
• Small object struggle.
• Computationally heavy (transformer).

Improvements:

• Deformable DETR — sparse attention.
• DINO-DETR — anchor-aware queries.
• RT-DETR — realtime, NMS-free.
• Co-DETR — collaborative head.

NMS-free trend implication:

• End-to-end training cleaner.
• Hyperparameter কম।
• Crowded scene natural handle.
• Still GPU expensive.

মূল উপলব্ধি: DETR's set prediction — detection-এর philosophical change. NMS heuristic-এর end ushering. Future direction clear.