Experimental evaluation of a generative probabilistic image retrieval model on 'easy' data

We present evaluation results of a generative probabilistic image retrieval model using `easy data'. Previous research into our model's retrieval effectiveness has used the test collection developed at TREC's Video Track, but as discussed in detail in [WeVr:SIGIR:03], its search task has been too difficult to measure actual performance of the retrieval model. The `easy data' experiments presented here evaluate our model under varying model parameters on the Corel set. The Corel data set is relatively easy because images are nicely grouped into coherent themes, the within theme similarity is high and the across theme similarity relatively low. These properties make Corel a nice vehicle for testing, presenting or selling new content based retrieval techniques and models. In contrast to the TREC data, the Corel collection gives statistically significant differences between varying experimental conditions, so we get more insight in the model's behaviour. We then discuss at length the limitations of the results obtained using this data set, comparing the experiments performed here to those on the TREC data.