Navigating Data Heterogeneity in Federated Learning: A Semi-Supervised Federated Object Detection
Authors: Taehyeon Kim, Eric Lin, Junu Lee, Christian Lau, Vaikkunth Mugunthan
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive validation on prominent autonomous driving datasets (BDD100K, Cityscapes, and SODA10M) attests to the efficacy of our approach, demonstrating state-of-the-art results. |
| Researcher Affiliation | Collaboration | Taehyeon Kim1 Eric Lin2 Junu Lee3 Christian Lau2 Vaikkunth Mugunthan2 1KAIST 2Dynamo FL 3The Wharton School potter32@kaist.ac.kr |
| Pseudocode | Yes | Algorithm 1: Fed STO Algorithm within the SSFOD Framework |
| Open Source Code | No | The paper does not contain an unambiguous statement or a direct link indicating that the authors are releasing the source code for the methodology described in the paper. While it mentions YOLOv5, that is a third-party tool. |
| Open Datasets | Yes | Extensive validation on prominent autonomous driving datasets (BDD100K, Cityscapes, and SODA10M) attests to the efficacy of our approach, demonstrating state-of-the-art results. (BDD100K [41], Cityscapes [4], SODA10M [9]) |
| Dataset Splits | Yes | For our studies, we employ the package, encompassing fine annotations for 3,475 images in the training and validation sets, and dummy annotations for the test set with 1,525 images. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU models, or cloud instance types) used for running its experiments. |
| Software Dependencies | No | The paper mentions "YOLOv5 Large model architecture" but does not specify version numbers for other key software components, libraries, or programming languages used. |
| Experiment Setup | Yes | Our training regimen spans 300 rounds: 50 rounds of warm-up, 100 rounds of pretraining (T1), and 150 rounds of orthogonal enhancement (T2). We use the YOLOv5 Large model architecture with Mosaic, left-right flip, large scale jittering, graying, Gaussian blur, cutout, and color space conversion augmentations. A constant learning rate of 0.01 was maintained. Binary sigmoid functions determined objectiveness and class probability with a balance ratio of 0.3 for class, 0.7 for object, and an anchor threshold of 4.0. The ignore threshold ranged from 0.1 to 0.6, with an Non-Maximum Suppression (NMS) confidence threshold of 0.1 and an Io U threshold of 0.65. We incorporate an exponential moving average (EMA) rate of 0.999 for stable model parameter representation. |