Scalable Video Genre Classification and Event Detection - Multimedia Database Retrieval: Technology and Applications - page 249

Fig. 9.2 Feature extraction and genre categorization framework using data parallelism and
bottom-up structure for codebook generation

9.2.2 Bottom-Up Codebook Generation

Local invariant features are chosen for homogenous feature extraction due to their
domain knowledge-free properties. The scale, rotation, and illumination invariant
properties make these descriptors good candidates in preserving the similarities
for semantic objects and events matching and detection. Global features, on the
other hand, rely on domain knowledge and have difficulties in robust concept
and event detection, especially in the presence of noise and occlusion [ 258 ].
Scale-invariant feature transform (SIFT), developed by Lowe [ 259 ], is selected