Self-Supervised Intrinsic Image Decomposition

Authors: Michael Janner, Jiajun Wu, Tejas D. Kulkarni, Ilker Yildirim, Josh Tenenbaum

NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method performs well on both intrinsic image decomposition and knowledge transfer.
Researcher Affiliation Collaboration Michael Janner MIT janner@mit.edu Jiajun Wu MIT jiajunwu@mit.edu Tejas D. Kulkarni Deep Mind tejasdkulkarni@gmail.com Ilker Yildirim MIT ilkery@mit.edu Joshua B. Tenenbaum MIT jbt@mit.edu
Pseudocode No The paper describes the model architecture and process in text but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing code, nor does it include a link to a code repository.
Open Datasets Yes The majority of data was generated from Shape Net [Chang et al., 2015] objects rendered in Blender. For the labeled datasets, the rendered composite images were accompanied by the object’s reflectance, a map of the surface normals at each point, and the parameters of the lamp used to light the scene. Surface normals are visualized by mapping the XYZ components of normals to appropriate RGB ranges. For the following supervised learning experiments, we used a dataset size of 40,000 images.
Dataset Splits No The paper mentions a dataset size of 40,000 images for supervised training and talks about 'test set' and 'unlabeled data' but does not provide explicit train/validation/test splits with percentages or sample counts to reproduce the data partitioning. No specific validation split is mentioned.
Hardware Specification No The paper does not provide specific hardware details (like GPU/CPU models or memory) used for running experiments.
Software Dependencies No The paper mentions tools like 'Blender' and 'Adam' but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes The encoder has 5 convolutional layers with {16, 32, 64, 128, 256} filters of size 3x3 and stride of 2. Batch normalization [Ioffe and Szegedy, 2015] and Re LU activation are applied after every convolutional layer. For both modes of learning, we optimize using Adam [Kingma and Ba, 2015]. During transfer, one half of a minibatch will consist of the unlabeled transfer data the other half will come from the labeled data.