Complex-Valued Autoencoders for Object Discovery
Authors: Sindy Löwe, Phillip Lippe, Maja Rudolph, Max Welling
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the Complex Auto Encoder on three grayscale datasets: 2Shapes, 3Shapes, and MNIST&Shape. For each of these datasets, we generate 50,000 training images and 10,000 images for validation and testing, respectively. All images contain 32 32 pixels. The 2Shapes dataset represents the easiest setting, with two randomly placed objects ( , ) in each image. The 3Shapes dataset contains a third randomly placed object ( ) per image. This creates a slightly more complex setting due to the higher object count, the two similar shapes ( , ), and stronger overlap between objects. Finally, the MNIST&Shape dataset combines an MNIST digit (Le Cun et al., 2010) and a randomly placed shape ( or ) in each image. This creates a challenging setting with more diverse objects. Finally, for evaluation, we generate pixel-accurate segmentation masks for all images. More details in Appendix C.4. |
| Researcher Affiliation | Collaboration | Sindy Löwe EMAIL Uv A-Bosch Delta Lab, University of Amsterdam; Phillip Lippe EMAIL QUVA Lab, University of Amsterdam; Maja Rudolph EMAIL Bosch Center for AI; Max Welling EMAIL Uv A-Bosch Delta Lab, University of Amsterdam |
| Pseudocode | No | The paper describes the model architecture and methodology using prose and mathematical equations but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/loeweX/Complex Auto Encoder. |
| Open Datasets | Yes | Finally, the MNIST&Shape dataset combines an MNIST digit (Le Cun et al., 2010) and a randomly placed shape ( or ) in each image. |
| Dataset Splits | Yes | For each of these datasets, we generate 50,000 training images and 10,000 images for validation and testing, respectively. |
| Hardware Specification | Yes | All experiments are implemented in Py Torch 1.10 (Paszke et al., 2019) and were run on a single Nvidia GTX 1080Ti. |
| Software Dependencies | Yes | All experiments are implemented in Py Torch 1.10 (Paszke et al., 2019) |
| Experiment Setup | Yes | We train the model using Adam (Kingma & Ba, 2015) and a batch-size of 64 for 10,000 100,000 steps depending on the dataset. Within the first 500 steps of training, we linearly warm up the learning rate (Goyal et al., 2017) to its final value of 1e 3. |