Semantic MapNet: Building Allocentric Semantic Maps and Representations from Egocentric Views

Authors: Vincent Cartillier, Zhile Ren, Neha Jain, Stefan Lee, Irfan Essa, Dhruv Batra964-972

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the task of semantic mapping in the Matterport3D dataset, SMNet significantly outperforms competitive baselines by 4.01 16.81% (absolute) on mean-Io U and 3.81 19.69% (absolute) on Boundary-F1 metrics. Moreover, we show how to use the neural episodic memories and spatio-semantic allocentric representations built by SMNet for subsequent tasks in the same space navigating to objects seen during the tour ( Find chair ) or answering questions about the space ( How many chairs did you see in the house? ).
Researcher Affiliation Collaboration Vincent Cartillier,1 Zhile Ren,1 Neha Jain,1 Stefan Lee,2 Irfan Essa,1,4 Dhruv Batra,1,3 1 Georgia Institute of Technology 2 Oregon State University 3 Facebook AI Research 4 Google Research {vcartillier3, zren80, nehaj, irfan, dbatra}@gatech.edu, leestef@oregonstate.edu
Pseudocode No The paper describes the architecture and modules of SMNet but does not include any pseudocode or algorithm blocks.
Open Source Code No The paper provides a "Project page: https://vincentcartillier.github.io/smnet.html", which is typically an overview/demo page and not explicitly a source code repository link as required by the prompt.
Open Datasets Yes We choose the Matterport3D dataset (Chang et al. 2017) with the Habitat simulator (Savva et al. 2019)... Matter Port3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.
Dataset Splits Yes Utilizing the same data split as (Wijmans et al. 2019), we keep 85 unique floors in our dataset: 61 for training, 7 for validation, and 17 for testing.
Hardware Specification Yes SMNet is trained end-to-end under cross-entropy loss using SGD with learning rate 1e 4, momentum 0.9, weight decay 4e 4, and batch size 8 across 8 Titan XPs.
Software Dependencies No The paper mentions software like Red Net, Habitat simulator, and SGD but does not provide specific version numbers for any of these or other key software components.
Experiment Setup Yes SMNet is trained end-to-end under cross-entropy loss using SGD with learning rate 1e 4, momentum 0.9, weight decay 4e 4, and batch size 8 across 8 Titan XPs. Training took 2-3 days. Back propagation is applied after 20 steps.