Differentiable Mathematical Programming for Object-Centric Representation Learning
Authors: Adeel Pervez, Phillip Lippe, Efstratios Gavves
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our results show that our approach is scalable and outperforms existing methods on object discovery tasks with textured scenes and objects. 6 EXPERIMENTS |
| Researcher Affiliation | Academia | Adeel Pervez QUVA Lab, Informatics Institute University of Amsterdam a.a.pervez@uva.nl Phillip Lippe QUVA Lab, Informatics Institute University of Amsterdam p.lippe@uva.nl Efstratios Gavves QUVA Lab, Informatics Institute University of Amsterdam e.gavves@uva.nl |
| Pseudocode | Yes | Algorithm 1 Feature k-Part Partitioning, Algorithm 2 Slots Computation, Algorithm 3 Matching slots for scene pairs. |
| Open Source Code | Yes | Code repository: https://github.com/alpz/graph-ocl |
| Open Datasets | Yes | We experiment with the Clevr (Johnson et al., 2017) and Clevr Tex datasets (Karazija et al., 2021). For Clevr we use the version with segmentation masks 1. https://github.com/deepmind/multi_object_datasets |
| Dataset Splits | Yes | Clevr Tex also provides two evaluation sets, OOD: an out-of-distribution dataset with textures and shapes not present in the training set and CAMO: a camouflaged version with the same texture used for the objects and background. Example reconstructions from the validation set can be seen in Fig 4. |
| Hardware Specification | Yes | We train using either a single A6000 GPU or 4 GTX 1080Ti GPUs for 1.5 days for Clevr Tex and on Clevr using a batch size of 64. |
| Software Dependencies | No | The paper mentions 'Cu Py (Okuta et al., 2017) which provides an interface to cu SPARSE (Nvidia, 2014)', but it does not specify explicit version numbers for these software dependencies, only citing the papers they were introduced in. |
| Experiment Setup | Yes | We train using either a single A6000 GPU or 4 GTX 1080Ti GPUs for 1.5 days for Clevr Tex and on Clevr using a batch size of 64. We use Adam with a learning rate of 4e-4. Training is generally stable and we did not require learning rate warm-up. Although not strictly necessary, we trained with an exponential learning rate decay. For these experiments we used 12 slots with each slot having a size of 64. We use a fixed softmax temperature of 0.1 for the combined slots and a temperature of 0.5 for the foreground slots. We set the regularization parameter γ = 0.5. |