Beyond Accuracy: A CNN Baseline and Failure Analysis for Landmark Classification on WikiScenes

Landmark classification remains a challenging task in computer vision due to high intra-class variation, long-tailed distributions, and real-world conditions such as occlusion and changes in viewpoint. This work introduces a systematic RGB-only formulation of the WikiScenes dataset, reframed as a 99-class benchmark that enables controlled evaluation of landmark recognition without relying on multimodal cues. We evaluate ResNet-18 and ResNet-50 across six distinct training regimes, including fine-tuning, training from scratch, and using frozen backbones. To address class imbalance, we apply weighted sampling and data augmentation, and assess performance using top-1 accuracy, micro-mAP*, and macro-mAP. Fine-tuned pretrained models consistently outperform other approaches. ResNet-18 achieves the highest micro-mAP of 0.8364 and macro-mAP of 0.7120, while ResNet-50 reaches the best top-1 accuracy at 84.3 percent. PCA embeddings and confusion matrices reveal that fine-tuned models produce compact and separable representations, whereas weaker training regimes result in diffuse and overlapping clusters. By establishing WikiScenes as a reproducible RGB-only benchmark and analyzing systematic failure modes, this study provides new insights into the limitations of image-only convolutional neural networks.