Object detection কী

এটি detection architecture design-এর central challenge. Multiple solutions evolved.

Approach 1: Dense grid prediction (YOLO):

• Image-কে $S \times S$ grid-এ ভাগ।
• Each cell predict $B$ box (typical 3-5).
• Total $S \times S \times B$ candidate output.
• Confidence threshold + NMS — final detection.
• Fixed output: $S \times S \times B \times (5 + C)$।

Approach 2: Region proposal (Faster R-CNN):

• RPN proposes ~2000 candidate region.
• Each candidate independent classify + refine.
• Per-image variable count detection.
• Fixed at proposal stage, variable at output.

Approach 3: Set prediction (DETR):

• Fixed $N$ "object queries" (e.g., 100).
• Each query — one slot in output (or "no object").
• Hungarian matching loss.
• Truly fixed output size.

Approach 4: Anchor-free (FCOS, CenterNet):

• Each pixel predict — center? size?
• Heatmap-style output.
• Local maxima → detection.

NMS — common deduplication:

• Multiple cell same object detect.
• NMS — highest confidence keep, IoU > 0.5 with high — discard.
• Variable output count emerge.

Trade-offs:

• Dense grid: dense candidate, NMS-dependent.
• Two-stage: accurate, slow.
• DETR: elegant set, slow convergence.
• Anchor-free: simpler, gaining popularity.

Loss function complexity:

• Each ground-truth box → assign to one prediction (anchor matching).
• Unassigned predictions → "background" class.
• Loss combination: classification + localization + objectness.

মূল উপলব্ধি: Variable output — neural network fundamental limitation. Detection-এর architectural creativity এই challenge-কে different ways-এ address. Modern trend — set prediction (DETR) elegant.

এটি detection community-র ongoing debate. Different metric different reality capture.

mAP@0.5 (Pascal VOC standard):

• Single IoU threshold.
• "Box approximately right" — counted.
• Lenient localization criterion.
• Pre-2017 standard.

mAP@0.5:0.95 (COCO):

• 10 IoU threshold (0.5, 0.55, ..., 0.95) average.
• Tight localization rewarded.
• Higher bar — harder to score high.
• Modern standard.

Numerical difference:

• Faster R-CNN: mAP@0.5 = 70%, mAP@0.5:0.95 = 35-40%.
• YOLOv8: mAP@0.5 = 65%, mAP@0.5:0.95 = 47%.
• "Better localization" matters.

Production metric choice:

• Self-driving: mAP@0.7 or higher — safety, tight box matters.
• Surveillance: mAP@0.5 sufficient — "presence detect" goal.
• Medical: mAP@0.7-0.9 — precise lesion border.
• Counting: mAP@0.5 — count more important than tightness.
• Robot grasping: mAP@0.9 — tight box critical.

Other relevant metrics:

• AR (Average Recall): max detection / proposal recall.
• AP-small/medium/large: size-stratified.
• F1 score at threshold: single confidence level.
• FPS / latency: realtime constraint.

Beyond traditional mAP:

• Open-vocabulary detection: CLIP-based — zero-shot class.
• Probability calibration: confidence reliability.
• Robustness: COCO-Stuff, OOD eval.

Caveats:

• mAP — single number, hides per-class variation.
• Class imbalance — rare class poorly score.
• "Hard negative" — confused background as object.
• Always look at confusion matrix, qualitative samples.

Bangladesh use case:

• Traffic detection — mAP@0.5 + size-specific.
• Crop disease — mAP@0.5 + class-balanced.
• Document text — IoU > 0.7 (tight letter box).

মূল উপলব্ধি: Single metric oversimplify. Production-এ multiple metric track + qualitative review. Numbers tell story, but visual inspection ground truth.

২০১৬-এর strict dichotomy ক্রমশ blur হচ্ছে। কিন্তু core trade-off relevant.

Historical context:

• Two-stage era (2014-2017): R-CNN → Fast R-CNN → Faster R-CNN. Accuracy lead.
• One-stage emergence (2016+): YOLO, SSD, RetinaNet — speed-focused.
• Modern: YOLOv8, RT-DETR — accuracy gap closed.

Two-stage strengths:

• Localization accuracy: RPN refine + RoI align — tight boxes.
• Small object: proposal at multiple scales.
• Custom objective: stage modification flexible.
• Used in: Mask R-CNN, Cascade R-CNN.

One-stage strengths:

• Speed: 30-150+ FPS.
• Mobile-friendly: lightweight backbones, quantizable.
• Simpler training: single loss.
• Used in: realtime application.

Modern landscape:

• YOLOv8 (2023): mAP 53.9% on COCO, 280 FPS.
• Faster R-CNN ResNet-50: mAP 41.0%, 17 FPS.
• RT-DETR: 53% mAP, 100 FPS.
• DINO-DETR: 63% mAP, slow.

Choice criteria:

• Realtime (>30 FPS): YOLOv8, RT-DETR.
• Maximum accuracy: DINO-DETR, Co-DETR.
• Mobile/edge: YOLOv8-nano, MobileDet.
• Instance segmentation: Mask2Former (DETR-style).
• Custom annotation type: two-stage flexibility.

DETR family — third paradigm:

• Set prediction, no anchor, no NMS.
• Traditionally slower convergence.
• DINO, Co-DETR — accuracy SOTA.
• RT-DETR — realtime competitive.

Bangladesh deployment:

• Mobile traffic app: YOLOv8-nano.
• Server CCTV: YOLOv8-medium.
• Medical research: Faster R-CNN + RX, accuracy priority.
• Edge IoT: YOLOv8-tiny, MobileDet.

Future trend:

• Distinction blurring.
• "Anchor-free + NMS-free" — DETR direction.
• Foundation model detection (Grounding DINO) — text prompt!

মূল উপলব্ধি: Old binary (one vs two stage) misleading. Modern question — speed vs accuracy + which architecture family + foundation model leverage.

NMS — detection-এর "necessary evil". Heuristic, deterministic, কিন্তু modern criticism.

NMS algorithm:

1. Confidence-এ sort detections (descending).
2. Highest confidence keep.
3. IoU > threshold-এ overlap-এর সব discard.
4. Repeat next highest confidence.
5. Continue till empty.

def nms(boxes, scores, iou_thresh=0.5):
    keep = []
    sorted_idx = scores.argsort(descending=True)
    while len(sorted_idx) > 0:
        idx = sorted_idx[0]
        keep.append(idx)
        ious = compute_iou(boxes[idx], boxes[sorted_idx[1:]])
        sorted_idx = sorted_idx[1:][ious <= iou_thresh]
    return keep

Why needed:

• Detector dense prediction — same object multiple time.
• Without NMS — output cluttered.
• Visualization inspires confusion.

NMS-এর problems:

• Hyperparameter: IoU threshold task-dependent.
• Crowded scene: two cars touching → one suppress.
• Inference cost: $O(N^2)$ box pairs.
• Non-differentiable: end-to-end training-এ obstacle.

Variants:

• Soft-NMS: linear/Gaussian decay instead of hard suppress.
• Weighted NMS: bbox merge instead of suppress.
• Adaptive NMS: threshold per-instance.
• Matrix NMS: parallelizable.

NMS-free detectors:

• DETR: set prediction — Hungarian matching ensure unique.
• Sparse R-CNN: learnable proposals.
• FCOS-style centerness: implicit duplicate suppress.

Modern hybrid:

• Most production system এখনো NMS use.
• YOLOv8: NMS.
• RT-DETR: NMS-free!
• Co-DETR: NMS-free with high accuracy.

Realtime impact:

• YOLOv8 GPU: NMS 1-2 ms.
• Mobile: NMS often biggest bottleneck post-conv.
• NMS-free — mobile-attractive.

Crowded scene — case study:

• Pedestrian dense crowd — overlap inevitable.
• Standard NMS → many missed.
• Soft-NMS → some recover.
• DETR-style → handles natively.

Bangladesh context:

• Dhaka traffic — extremely crowded.
• Standard YOLO + NMS — overlap rickshaw + CNG miss.
• RT-DETR or specialized crowd detector preferable.

মূল উপলব্ধি: NMS — historical artifact. Modern direction NMS-free. কিন্তু practical deployment-এ এখনো ubiquitous. Tool change হলেও limitation conceptual.