Rotating Features for Object Discovery
Authors: Sindy Löwe, Phillip Lippe, Francesco Locatello, Max Welling
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate whether our proposed improvements enable distributed object-centric representations to scale from simple toy to real-world data. We begin by outlining the general settings common to all experiments, and then proceed to apply each proposed improvement to a distinct setting. This approach allows us to isolate the impact of each enhancement. Finally, we will highlight some advantageous properties of Rotating Features. Our code is publicly available at github.com/loewe X/Rotating Features. |
| Researcher Affiliation | Academia | Sindy L owe Phillip Lippe Francesco Locatello AMLab QUVA Lab Institute of Science and University of Amsterdam University of Amsterdam Technology Austria (ISTA) Max Welling AMLab University of Amsterdam |
| Pseudocode | No | The paper describes the methods using mathematical equations and textual explanations, but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is publicly available at github.com/loewe X/Rotating Features. |
| Open Datasets | Yes | 4Shapes The 4Shapes dataset comprises grayscale images of dimensions 32 32... The dataset consists of 50,000 images in the train set, and 10,000 images for the validation and test sets, respectively. Pascal VOC The Pascal VOC dataset [19] is a widely-used dataset for object detection and semantic segmentation. Food Seg103 The Food Seg103 dataset [85] serves as a benchmark for food image segmentation... |
| Dataset Splits | Yes | 4Shapes The 4Shapes dataset comprises grayscale images of dimensions 32 32... The dataset consists of 50,000 images in the train set, and 10,000 images for the validation and test sets, respectively. We evaluate the performance of Rotating Features on the official segmentation validation set, containing 1,449 images. |
| Hardware Specification | Yes | Our experiments are implemented in Py Torch [54] and run on a single Nvidia GTX 1080Ti. |
| Software Dependencies | No | The paper mentions software packages like PyTorch, scikit-learn, scikit-image, and timm, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | General Setup: We implement Rotating Features within a convolutional autoencoding architecture. Each model is trained with a batch-size of 64 for 10,000 to 100,000 steps, depending on the dataset, using the Adam optimizer [40]. Hyperparameters All hyperparameters utilized for models trained directly on raw input images are detailed in Table 3, while those for models trained on DINO features can be found in Table 4. |