qhjqhj00/research-skills-pool/aid-aerial-scene-classification-eval

aid-aerial-scene-classification-eval

AID: A Benchmark Dataset for Performance Evaluation of Aerial Scene Classification — Gui-Song Xia et al. (2016) (arXiv:1608.05167, 2016)

What this evaluates

Evaluates the ability of computer vision models to classify aerial imagery into distinct scene categories. It probes robustness to high intra-class diversity and low inter-class similarity in remote sensing data.

Datasets

• AID — total ?; splits: train (-1), test (-1)

Metrics

• accuracy (primary) — range: [0, 1]
• Standard classification accuracy: the proportion of correctly predicted scene categories out of the total number of test images.

Input / output format

Input: RGB aerial images representing diverse scene categories.

Output: A single discrete class label corresponding to the aerial scene category.

Scoring recipe

correct = 0
for img, true_label in test_set:
    pred_label = model.predict(img)
    if pred_label == true_label:
        correct += 1
accuracy = correct / len(test_set)

Common pitfalls

• Dictionary size (K) and number of latent topics (T) for mid-level methods (BoVW, pLSA, LDA) are critical hyperparameters that drastically affect feature dimensionality and performance, yet are often reported without ablation.
• High-level methods use pre-trained ImageNet weights without fine-tuning on AID, which may underestimate their potential compared to fine-tuned or fully trained variants.
• Intra-class diversity is high and inter-class similarity is low, making standard Euclidean distance or simple pooling less effective without spatial pyramid matching or locality constraints.

Evidence (verbatim from paper)

In this section, we evaluate different aerial scene classification methods with low-, mid- and high-level scene descriptions reviewed previously. ... shows a significant improvement on the accuracies

Citation

@misc{xia2016aid,
  title={AID: A Benchmark Dataset for Performance Evaluation of Aerial Scene Classification},
  author={Gui-Song Xia et al. (2016)},
  year={2016},
  note={arXiv:1608.05167}
}

• arXiv: 1608.05167