You can test Accio! with two different image sets. SIVAL is an image set in which a 25 different objects are placed among a variety of different backgrounds. Also, the search can be for an object from the background (e.g. a fire hydrant) as opposed to one of the 25 different categories. The second image set is natural scenes from Corel's sunset, field, flower, mountain and waterfall categories.
NOTE: Currently it takes about 45 seconds for a search. Only a couple of seconds are from the learning algorithm. We are working on modifying the scripts to remove the extra delay. Please be patient until we speed things up.
In Step 1 you can either select the SIVL database (25 objects) or the Natural Scenes database. Then select a segmentation method. (In some cases, currently just one choice is currently provided.) Using more segments generally improves the quality of the results at the cost of a higher computational costs. Next you can select if neighbors are to be used. Generally, using neighbors gives better results, but is computationally more expensive. The option of supplying your own primer image is not yet available. That option will be added in the near future.
In Step 2 you can select a category and then either see a random selection of images from that category, or all images. This step is designed to just start the search. Clicking on an image moves it between not selected (white frame), positive (green frame) and negative (red frame). You can just select one image as positive, or create a larger training set. Eventually, we will allow you to provide your own image to begin the search, but that function is not currently available.
Step 3 is where relevance feedback can be applied as many steps as needed. It will show the top ranked 25 images (in row major order) from the most recent search performed. When you move the mouse over an image, the segmented image is shown. Also, for each image you can view a highlight showing a visualization of which segments within the image were most important to the ranking, and giving the image filename. (The filename is not used in ranking the images.) As in Step 2, clicking on an image moves it between not selected (white frame), positive (green frame) and negative (red frame). After each search you can augment your search (adding the marked positive and negative images to the training data) or start a new search (using the marked positive and negative images as a new training set). You can also return to Step 1 and begin again.