90 likes | 257 Views
TI’s HOG implementation details. BOP setup and assumptions. ROI size is 240 x 160 ( entire down-sampled image) Block sizes within ROI are 64x96 Blocks are further divided into Cells, E ach of 12x12 Each of 16x16 Each of 24x24 Each of 32x32 Assumptions: Each cell gives a feature
E N D
BOP setup and assumptions • ROI size is 240 x 160 (entire down-sampled image) • Block sizes within ROI are 64x96 • Blocks are further divided into Cells, • Each of 12x12 • Each of 16x16 • Each of 24x24 • Each of 32x32 • Assumptions: • Each cell gives a feature • Smaller cells give better features, so fewer larger cells are considered in the blocks • As blocks are part of larger ROI, looks like pixels from adjacent ROIs are considered in some cases where cells cross block boundaries
Each ROI is divided into Blocks as shown, skip amount 20 pixels
Each block is divided into Cells (each of size 12x12), skip amount 8 pixels
Each block is divided into Cells (each of size 16x16), skip amount 8 pixels
Each block is divided into Cells (each of size 24x24), skip amount 8 pixels
Each block is divided into Cells (each of size 32x32), skip amount 8 pixels
Notes • Note that for smaller cell sizes we compute more features • For larger cell sizes, fewer features are computed • I think, this is because smaller cell sizes give better feature. • Also, some cell blocks cross the block boundary, this is also probably ignored as the block is part of a larger ROI and the pixels are still available and valid.
Notes • For each block we get • 756+756+432+180 = 2124 features • There are 36 such blocks • Classifier will use these features and do binary classification with 51 coefficients • All 36 blocks will be considered to decide whether an object of interest is present in the field of view or not.