Ssd Voc2007 Eval

ssd-voc2007-eval

SSD: Single Shot MultiBox Detector — Wei Liu et al. (2015) (arXiv:1512.02325, 2015)

What this evaluates

Evaluates real-time object detection capability by predicting bounding boxes and class scores directly from multi-scale feature maps, eliminating traditional proposal generation steps. It measures how well the model localizes and classifies objects across varying scales and aspect ratios under strict latency constraints.

Datasets

• PASCAL VOC2007 — total 4952; splits: test (4952)

Metrics

• mAP (primary) — range: percent
  • Mean Average Precision averaged over 20 object categories. Computed as the area under the precision-recall curve for each class, then averaged.

Input / output format

Input: RGB image resized to 300x300 or 512x512 pixels

Output: List of predicted bounding boxes with associated class labels and confidence scores

Scoring recipe

def compute_mAP(preds, gts):
    ap_scores = []
    for cls in range(20):
        cls_preds = sorted([p for p in preds if p['class'] == cls], key=lambda x: x['score'], reverse=True)
        cls_gts = [g for g in gts if g['class'] == cls]
        tp, fp, matched = 0, 0, set()
        for p in cls_preds:
            best_iou, best_gt = 0, None
            for g in cls_gts:
                if g['id'] not in matched:
                    iou = calc_iou(p['bbox'], g['bbox'])
                    if iou > best_iou: best_iou, best_gt = iou, g
            if best_iou >= 0.5: tp += 1; matched.add(best_gt['id'])
            else: fp += 1
        precisions = [tp / (tp + fp + 1e-9) for _ in range(len(cls_preds))]
        recalls = [tp / max(len(cls_gts), 1) for _ in range(len(cls_preds))]
        ap_scores.append(trapezoidal(precisions, recalls))
    return sum(ap_scores) / 20

Common pitfalls

• Performance degrades significantly on small objects due to limited information in top feature layers.
• High confusion rates between visually similar categories (e.g., animals) because the model shares location predictions across classes.

Evidence (verbatim from paper)

Table 1 shows that our low resolution SSD300 model is already more accurate than Fast R-CNN. When we train SSD on a larger 512 × 512 input image, it is even more accurate, surpassing Faster R-CNN by 1.7% mAP.

Citation

@misc{liu2015ssd,
  title={SSD: Single Shot MultiBox Detector},
  author={Wei Liu et al. (2015)},
  year={2015},
  note={arXiv:1512.02325}
}

• arXiv: 1512.02325