Retrievr

Retrievr is an experimental service which lets you search and explore Flickr images by drawing rough sketches or uploading images. Finding images on Flickr is mostly textual, for example through tags and keywords, or social, for example, through friends or groups. Retriever, on the other side, is using visual images as queries. Retrievr brings back 20 results for each query, all of them appear as thumbnail in one page. By clicking the thumbnail, you go to the original copy of it in Flickr and get more information such as author, tags, and comments.

Here is the result page of my try of drawing a red heart: http://labs.systemone.at/retrievr/#sketchName=2008-02-13-17-55-25-840115.12. To me, it seems like images retrieved by Retriever were of similar color (the red chair), or of similar shape (the baby face), or of both (the red heart) to the sketching query. Retrievr search images matching the most pronounced shapes and slabs of colors.

Later after playing with it for a while, I found other interesting aspects of Retrievr.

(1) Retrievr retrieves different set of images when the position of the sketch on the sketching board changes. Instead of putting the heart in the middle of the sketching board, I drew another one in the top right corner of the sketching board. Retrievr gave me a totally different set of images and the right corners of the images matched my sketch. http://labs.systemone.at/retrievr/ - sketchName=2008-02-13-17-55-25-840115.30

(2) Retrievr cannot identify letters. I wrote a red “A” and got back no image containing the letter A. http://labs.systemone.at/retrievr/?sketchName=2008-02-13-17-55-25-840115.3#sketchName=2008-02-15-20-15-47-829302.2

(3) There were always images taking position in the result list but were not actually available (instead of the real thumbnail of the image, it shows an icon saying this photo is currently unavailable.) It makes me doubt if Retrievr does a real time search in Flickr’s database. One possible explanation is it might suck images, possibly only part not all, from the Flickr, process them technically to improve the search performance, and then store them in their own database. When it tries to link back to the original image in the result page, it happens that the image has been taken off by the author.

Retrievr works pretty well. It always brings me several matching images out of the twenty every time. I here have some thoughts which may improve the interaction between the system and the user.

(1) Is there any possibility that Retrievr could retrieve more images for each query? Currently, Retrievr only brings up 20 images and I think it should be able to get more from Flickr, which is such a huge collection of images, to satisfy users with high recall expectation.

(2) Is there any possibility that Retrievr could cluster the retrieved images, for example by shape, by color?

(3) Is there any possibility that Retrievr could offer some interaction letting users with high precise expectation give feedbacks on the first try of the search and improve the search results by doing a second one with users’ feedback in mind?

(4) Is there any possibility that Retrievr could get rid of those images which are not available at the moment of the user conducts the search. Why images which could not been seen take up positions in the result pages?