Relational Proxies: Emergent Relationships as Fine-Grained Discriminators
Authors: ABHRA CHAUDHURI, Massimiliano Mancini, Zeynep Akata, Anjan Dutta
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We design Relational Proxies based on our theoretical findings and evaluate it on seven challenging fine-grained benchmark datasets and achieve state-of-the-art results on all of them, surpassing the performance of all existing works with a margin exceeding 4% in some cases. We also experimentally validate our theory on fine-grained distinguishability and obtain consistent results across multiple benchmarks. |
| Researcher Affiliation | Academia | 1 University of Exeter 2 University of Tübingen 3 MPI for Informatics 4 MPI for Intelligent Systems 5 University of Surrey |
| Pseudocode | No | The paper describes its methods but does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Implementation is available at https://github.com/abhrac/relational-proxies. |
| Open Datasets | Yes | We evaluate our model on the four most common fine-grained visual categorization benchmarks (number of classes and train/test splits respectively in brackets): FGVC Aircraft [26], Stanford Cars [20], CUB [41], and NA Birds [39]. For large scale benchmark evaluation, we choose the i Naturalist 2017 dataset which consists of 13 super-categories that have been split into a total of 5089 fine-grained categories with 675,170 training and 182,707 test images. We also perform experiments on two challenging datasets of the cultivar domain that offer very low inter-class variations, namely Cotton Cultivar [46] (80 | 240/240) and Soy Cultivar [46] (200 | 600/600). |
| Dataset Splits | No | The paper provides train/test splits for the datasets but does not explicitly mention a separate validation split with specific sizes or percentages. |
| Hardware Specification | Yes | We implement our Relational Proxy model using the Py Torch [32] deep learning framework, on an Ubuntu 20.04 workstation with a single NVIDIA Ge Force RTX 3090 GPU, an 8-core Intel Xeon processor and 32 GBs of RAM. |
| Software Dependencies | No | The paper mentions using 'Py Torch [32] deep learning framework' but does not specify a version number for PyTorch or other software dependencies like Python or CUDA. |
| Experiment Setup | Yes | We use Res Net50 [14] pretrained on Image Net [9] as the backbone of our relation-agnostic encoder f. In Sec. 1.3 of the supplementary, we also provide evaluations using VGG-16 [36] to show that the performance gains achieved by our model do not depend on the specific backbone. We train our full Relational Proxy model end-to-end for 200 epochs using the stochastic gradient descent optimizer with an initial learning rate of 0.001 (decayed by a factor of 0.1 every 50 epochs), a momentum of 0.9, and a weight decay of 10 4. |