----------------------------------------------------------------- --- A. Results ----------------------------------------------------------------- The directory "Results" contains the training/testing results of MLN-EC and l-CRF methods for both tasks. The target composite events (CE) are the "meeting" and "moving" activities. TASK_I: ---------- For MLN-EC, both CE are trained using the Alchemy Markov Logic Networks and the M3LNs software packages (both can be found in http://alchemy.cs.washington.edu). All experiments are performed using 10-fold cross-validation on the CAVIAR dataset (http://homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/). MLN-EC is trained using either the Diagonal Newton method and Max-Margin. Each sub-directory contains the following configurations: 1. HI: The initiation/termination rules are soft-constrained, while the inertia rules remain hard-constrained. 2. SI: The initiation/termination rules, as well as the inertia rules are soft-constrained.-constrained. 3. SI_h: The initiation/termination rules are soft-constrained. From the inertia rules only the inertia of HoldsAt is soft-constrained. Each configuration contains the following files for both CE (i.e. meeting and moving): * Source KB (Move_src*.mln or Meet_src*.mln), in which the weight configurations (soft/hard) are set appropriately according to the current configuration (HI, SI or SI_h). * Training KB in clausal form (move.mln or meet.mln) that are generated using the 'compile.sh' script (located inside the directory) * Inside fold directories (meet/fold_*/ or move/fold_*): - The *_trained.mln is the resulting trained MLN file - The training directory contains the training database files (*.db) - The testing directory contains the testing input evidence files (*.db) and the testing result files (*.result) In the case of Diagonal Newton, the results are from marginal inference. In the case of Max-Margin the results are from MAP inference. Usage examples: All training KB files (i.e. meet.mln and move.mln) are generated using the compile.sh, which is located inside each training configuration, e.g. HI/meeting/compile.sh. This Bash script file uses the LoMRF library (can be found in www.iit.demokritos.gr/~anskarl/pub/LoMRF) in order to convert the source KB (e.g. HI/meeting/Meet_src-HI.mln) into the correct form for training, i.e. weights configuration, correct clausal form and replacement of functions with utility predicates (e.g. meet(p1,p2) is replaced by the predicate ReturnValueOfMeet(ret_val,p1,p2)). Example training settings for the Alchemy's learnwts executable: e.g. for the meeting CE, fold 0: learnwts -d -noAddUnitClauses -dNewton -dNumIter 500 \ -i HI/meeting/meet.mln \ -o HI/meeting/meet/fold_0/meet_trained.mln \ -t HI/meeting/meet/fold_0/training/training.fold_0.db \ -ne HoldsAt \ -cw Close,Next,Happens,OrientationMove,StartTime,ReturnValueOfactive,ReturnValueOfmeet,ReturnValueOfenter,ReturnValueOfrunning,ReturnValueOfwalking,ReturnValueOfinactive,ReturnValueOfexit e.g. for the moving CE, fold 0: learnwts -d -noAddUnitClauses -dNewton -dNumIter 500 \ -i HI/moving/move.mln \ -o HI/moving/move/fold_0/move_trained.mln \ -t HI/moving/move/fold_0/training/training.fold_0.db \ -ne HoldsAt \ -cw Close,Next,Happens,OrientationMove,StartTime,ReturnValueOfactive,ReturnValueOfmove,ReturnValueOfenter,ReturnValueOfrunning,ReturnValueOfwalking,ReturnValueOfinactive,ReturnValueOfexit Example testing settings for the Alchemy's infer executable: e.g. for the meeting CE, fold 0: infer -ms \ -i HI/meeting/meet/fold_0/meet_trained.mln \ -e HI/meeting/meet/fold_0/testing/21-Meet_WalkSplit.id1_id3.db \ -r HI/meeting/meet/fold_0/testing/21-Meet_WalkSplit.id1_id3.result \ -q HoldsAt \ -cw Close,Next,Happens,OrientationMove,StartTime,ReturnValueOfactive,ReturnValueOfmeet,ReturnValueOfenter,ReturnValueOfrunning,ReturnValueOfwalking,ReturnValueOfinactive,ReturnValueOfexit e.g. for the moving CE, fold 0: infer -ms \ -i HI/moving/move/fold_0/move_trained.mln \ -e HI/moving/move/fold_0/testing/21-Meet_WalkSplit.id1_id3.db \ -r HI/moving/move/fold_0/testing/21-Meet_WalkSplit.id1_id3.result \ -q HoldsAt \ -cw Close,Next,Happens,OrientationMove,StartTime,ReturnValueOfactive,ReturnValueOfmove,ReturnValueOfenter,ReturnValueOfrunning,ReturnValueOfwalking,ReturnValueOfinactive,ReturnValueOfexit TASK_II: ---------- The aim of the second task is to examine the recognition of CEs in the two probabilistic models, under situations where observations are incomplete. The incomplete test sequences are generated artificially by erasing successive input SDEs and spatial constraints at random starting points. We have generated variants of incomplete test sequences, containing random blank intervals of 5, 10, 15 and 20 time-point duration. The starting time-points of the blank intervals are chosen randomly with probability 0.01, drawn from a uniform distribution. Since the target CEs require the interaction of two persons, erasing events involving a single person cannot affect the performance of the recognition methods that we compare. Therefore, blank intervals are created only from time-points where both persons are involved in some SDEs. This process of artificially generating incomplete test sequences is repeated five times. ----------------------------------------------------------------- --- B. KB Compilation Examples ----------------------------------------------------------------- This directory contains some demo KB compilation examples for both "meeting" and "moving" CE, by using the LoMRF library (can be found in www.iit.demokritos.gr/~anskarl/pub/LoMRF). Further details can be found inside the comments of *.mln and compile.sh files of this directory.