Solving Partial Assignment Problems using Random Clique Complexes

Authors: Charu Sharma, Deepak Nathani, Manohar Kaul

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we present a comprehensive empirical study that compares our method s matching accuracy to that of a diverse set of matching approaches (Zhou & De la Torre, 2016; Zhou & De la Torre, 2013; Cho et al., 2010; Feizi et al., 2016; Leordeanu & Hebert, 2005; Cour et al., 2007; Pachauri et al., 2013; Gold & Rangarajan, 1996; Kuhn, 1955; Leordeanu et al., 2009; Zass & Shashua, 2008; Li et al., 2013; Duchenne et al., 2011). We conducted our experiments on both synthetic and well-known hard real-world datasets that span across affine/non-affine transformations, severe occlusions, and clutter. Our study reveals much better accuracy for the popular datasets against several of the state-of-the-art matching methods.
Researcher Affiliation Academia 1Department of Computer Science & Engineering, Indian Institute of Technology Hyderabad, Hyderabad, India.
Pseudocode Yes Algorithm 1 Matching Random Clique Complexes Input: X(G) = {G(k,l)}h k=0 and X(G ) = {G (k,l)}h k=0 1: for k = h . . . 0 do 2: Let M, M be the total number of (k + 1)-cliques in G(k,l) and G (k,l), respectively 3: L := {c(k) i }M 1 i=0 # list of barycenters 4: for i = 0 . . . M 1 do 5: Ni := Ni n g(k,l) (x,:) | x = i, g(k,l) (x,y) = 0 o 6: N := N {Ni} # clique neighborhoods 7: end for 8: for i = 0 . . . M 1 do 9: αi := [α1, . . . , αM 1]T 10: α := α {αi} # affine weight vectors 11: end for 12: Repeat steps 3 11 on G (k,l) for L , N and α . 13: Build cost matrix C(k) from weights vectors α, α 14: X k := Kuhn-Munkres (G(k,l), G (k,l), C(k)) 15: end for Return: {X 0, . . . , X h} # set of permutation matrices
Open Source Code No No explicit statement or link indicating that the source code for *this paper's* methodology is publicly available.
Open Datasets Yes We took two real-world datasets, i.e., Books and Building (Pachauri et al., 2013)
Dataset Splits Yes we uniformly sample frames (at 20% and 40%) and perform affine transformations on the selected frames to distort them. and we omit 2, 4, 6, 8, and 10 (6.66%, 13.33%, 20%, 26.66%, and 33.33%) points out of total House landmark points (i.e., 30 points) from 40% (Figure 4) of frame sequences randomly.
Hardware Specification No No specific hardware details (such as GPU/CPU models, memory, or cloud resources) were mentioned for running experiments.
Software Dependencies No No specific ancillary software dependencies (e.g., library or solver names with version numbers) are mentioned in the paper.
Experiment Setup Yes We set p = 0.7 and k = 7 as nearest neighbors to get the correct matchings.