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।