Why does segmentation in IMAGINE Objective show tiling artifacts that introduce seamlines along four quarter tiles?
Some segments are continuous across the quarter tile seamline and some segments have borders along the seamline.
Segmentation is a global operation where all pixels/segments need to be in memory to be computationally efficient. A tiling mechanism is implemented to process larger images. The algorithm tries to process a tile and, if it will not fit into memory, it quarters into for sub-tiles and recursively continues quartering tiles until it finds sub-tiles that can fit into memory. It then segments the four quarter tiles one at a time and then rolls the resulting quarter tile segments into the parent tile. The parent tile continues merging segments which can cause segments to merge across the tile boundary reducing the tile artifacts.
This approach help reduce tile boundary artifacts but it does not always eliminate them completely. For larger images there will sometimes be some tile boundary artifacts especially along the last four quarter tiles in the process (which correspond to the four quarter tiles of the entire input image). The segmentation parameters, particularly setting a large Minimum Segment Size, may contribute to boundary artifacts.
Furthermore, it is possible that these tile boundary artifacts can cause errors downstream during object classification but it depends on which object cues are used in the classification. Cues that are based on the object geometry, like area, circularity, etc., may be classified erroneously. But for cues that are not geometric in nature, such as, zonal mean, texture, EDA, etc., the tile boundary artifacts will not cause classification errors.
Regarding the size of input image, there is no hard limit. But the larger the image, the more likely there will be tile artifacts. Segmentation will work best in a 64 bit Windows OS version with a minimum of 4 GB memory.
Significant improvements are made in IMAGINE 2014 for FLS Segmentation. It is a true 64 bit process. The segmentation memory usage will scale to the computer’s available RAM. Larger images will be able to be processed more efficiently with fewer tile boundary artifacts. Users can run on systems that have 128 GB RAM, for example, and the segmentation process will not have to break into nearly as many tiles, and thereby, reduce these tile artifact boundaries.