2.5 Problems/Solutions

One of the main issues dealt with was that of loading time and more generally, computer response time.To save processing time and to improve response time, two versions of the database were created, onefor each treatment of thumbnail size. Thus, all computer time spent was devoted to redrawing and loadingand none was wasted on resizing. The disk space occupied by 176 images for the 5x5 grid (100x100) was about 2.5MB. Similarly, the space occupied by theimages used for the 3x3 grid (170x170) was over 5MB. This implies that all of the data must be in RAM in order to minimize theresponse time. While all the computers had enough installed RAM, since most used virtual memory, occasionally parts of the data needed tobe accessed was sometimes paged to disk in a non-deterministic manner. However, of greater concern was that due to a lackof resources, it was impossible to ensure that every subject was tested on the exact same machine. Hence, there was an unnecesaryvariability in response time due to the range of computers used. However, this variability was on the order of millisecondsand did not appear to have a significant impact on user performance times.

The original images in the Coolidge collection were much larger than the thumbnails used. The method used to resizethe images was using the Image Size function in Adobe Photoshop 4.0, with bicubic interpolation. This algorithmprovides the smoothest image downsampling, however it tends to blur details somewhat when a given image is resized to be much smaller than the original.Because of this side effect, the 100x100 images may have been perhaps less detailed than they could have been.However, this method was used consistently for all the images across all thumbnail sizes.

All of the images in the database had varying proportions. To deal with this problem, when images were resized, theiroriginal proportions were retained and the longest dimension of any given picture was fit to occupy the entire length ofthe appropriate side of the square. This implied that for any non-square image, parts of the thumbnail would be empty. The empty space wasfilled with an arbitrarily chosen black color. While this meant that images whose proportions were closer to that of a squarewould appear larger than those less so, nearly all of the images in the database had a proportion of either about 4/3 or 3/4, whichassured that no image would appear significantly smaller or larger than the others.

As for the issue of the learning effect of subjects (since each had six tasks on the same database), the large sizeof the database and the high similaritry of the images reduced the chance of learning occuring. Additionally, betweeneach task the order of images was randomized. Some subjects commented on the fact that the database was so large thatthey didn't know if the same set of images was being used between tasks. A few even mentioned being surprised that even afterlooking though the database many times, they did not remember previously seeing target images that matched the text descriptionfor subsequent tasks.