Learning Relational Event Models from Video
Authors: Krishna S. R. Dubba, Anthony G. Cohn, David C. Hogg, Mehul Bhatt, Frank Dylla
JAIR 2015 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present an evaluation of remind, as well as the extension presented in Section 7. For the experiments, we used two real world video datasets that are different from each other in many aspects. ... The experimental results on several hours of video data from two challenging real world domains (an airport domain and a physical action verbs domain) suggest that the techniques are suitable to real world scenarios. The results of our experiments are summarised in Table 1. |
| Researcher Affiliation | Academia | Krishna S. R. Dubba EMAIL Anthony G. Cohn EMAIL David C. Hogg EMAIL School of Computing, University of Leeds, Leeds, UK. LS2 9JT Mehul Bhatt EMAIL Frank Dylla EMAIL Cognitive Systems, SFB/TR 8 Spatial Cognition University of Bremen, Bremen 28334, Germany |
| Pseudocode | Yes | Algorithm 1 Interleaved Induction and Abduction algorithm (IIA) procedure IIA(E+, E , B) training sets and background knowledge (includes spatial theory) H while E+ = do Rule Induce(B, E+, E ) H H S{Rule} E+ E+ E+ Rule Abduce(B, H, E+) E+ E+ E+ end while return H Learned theory end procedure |
| Open Source Code | No | remind is implemented in Python and for speed, some modules are implemented in Cython; SWI-Prolog is used as the underlying Prolog engine for storing and querying relational facts and background knowledge. Available on request from the first author and will be made public in the near future. |
| Open Datasets | Yes | The Action Verbs dataset is a corpus of video vignettes (Fig.9) that portray motion verbs such as approach, exchange, jump, collide, etc. enacted in natural environments like parks,... This dataset (Mind s Eye Year 1 recognition task videos) is provided by DARPA and publicly available from http://www.visint.org/datasets |
| Dataset Splits | Yes | For supervised learning we need positive and preferably negative example instances of events. In the airport domain, the temporal extent of the events is provided by individuals who have expertise in the IATA protocols and apron activities, by specifying the start and end frame numbers of the event instance in that video. ... We followed the standard leave-one-out methodology for testing performance in the airport domain. All turn-arounds except one are used for training and the remaining one is used as a test case. This process is iterated until each turn-around is used as the test case exactly once. ... 10-fold cross-validation was used for evaluation for verbs dataset. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments. It mentions software like Python, Cython, and SWI-Prolog, but no hardware specifics. |
| Software Dependencies | No | remind is implemented in Python and for speed, some modules are implemented in Cython; SWI-Prolog is used as the underlying Prolog engine for storing and querying relational facts and background knowledge. |
| Experiment Setup | Yes | score(H) = γ p (ϱ n + l + v) γ = weight to positive examples p = number of positive examples covered ϱ = weight to answer substitutions in negative examples l = length of the hypothesis v = number of distinct variables in the hypothesis ... The cost of abducing fijrk is calculated as follows: ι, if there exists a fijrm in E+ p such that δk and δm are disjoint κ, if there exists a fijsm in E+ p such that δk and δm are disjoint and r = s µ n, where n is the number of hypothetical objects (objects not in O) in fijrk where ι < κ < µ. ... The algorithm makes use of domain-dependent and domain-independent constraints on the structure of the hypothesis. The constraints the algorithm uses such as restrictions on the hypothesis length and the number of variables in the hypothesis etc. are domain-independent structural constraints as they do not depend on the predicates used or any domain knowledge. The following are the two domain-dependent constraints that reduce the search space and time thereby making the learning process more efficient. Upper bounds on the number of atoms in the body of a rule. |