This study evaluates seven prominent SIFT implementations for feature detection in Wide Area Motion Imagery (WAMI): Lowe's archived code, VLFeat, OpenCV, SIFT anatomy, CudaSIFT, SiftGPU, and PopSift. We use spatio-temporal patch animations, termed ThumbTracks, to assess each method's performance in terms of jitter, wandering, and track switches. Additionally, we analyze the clustering of SIFT descriptors using t-distributed stochastic neighbor embeddings. Our results reveal significant variations in the performance of different SIFT variants, with implications for their suitability in various WAMI applications. We provide recommendations for selecting the most appropriate SIFT implementation based on feature stability, computational efficiency, and accuracy requirements.
|