6. September 2012 The version of trec_eval used in TRECVID was early on modified to adjust the calculation of average precision for a feature in the case where the total number of known true positives was greater than the maximum result size. This could easily happen since our tools for estimating total relevant for a given TRECVID feature are very rough. Without this adjustment it would be impossible for a run to achieve an average precision of 1, which should at least in theory be possible - the calculation of average precision would be artificially capped, and although equally for all runs, this would depress/distort the mean average precision. The program sample_eval.pl used up until now did not attempt to adjust inferred average precision AP in cases where the inferred number relevant for a feature is larger than the maximum result set size (2000 for semantic indexing). The program has now been modified for TRECVID 2012 to do that. Adjusted results for TRECVID 2011 and 2010 are provided in this directory. The modified version of sample_eval.pl is available in the TRECVID tools directory: http://www-nlpir.nist.gov/projects/trecvid/trecvid.tools/ Paul Over