Variational Feature Pyramid Networks
Authors: Panagiotis Dimitrakopoulos, Giorgos Sfikas, Christophoros Nikou
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this Section, we provide numerical results for the proposed method, in comparison to recent existing feature fusion networks. For our numerical analysis, we perform three different experiments. First, we evaluate our methods as a backbone network for detection, using (Ren et al., 2016) and instance segmentation, using (He et al., 2017) versus state-of-the-art backbone combinations. We carried out experiments to evaluate how the learned architecture of our network can adapt to different types of datasets, containing objects at various scales and sizes. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Engineering, University of Ioannina, Ioannina, Greece 2University of West Attica, Athens, Greece 3National Center for Scientific Research Demokritos , Athens, Greece. |
| Pseudocode | No | The paper provides mathematical formulations and textual descriptions of the model and its components, but it does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | No | The paper states: 'Our model implementations were based on the MMDetection open source project (Chen et al., 2019)', indicating they used an existing open-source project, not that they are providing their own code as open-source for the described methodology. |
| Open Datasets | Yes | Our main experiments are conducted on the large-scale detection benchmark COCO (Lin et al., 2014)... Plant Doc (Singh et al., 2019) contains (mostly) medium and large objects and the Cards dataset (Crawshaw, 2020) is comprised solely of small objects. |
| Dataset Splits | Yes | Following common practices (Lin et al., 2017b; Tian et al., 2019), we use the COCO trainval35k split (115K images) for training and the minival split (5K images) as validation. |
| Hardware Specification | Yes | We also gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan XP GPU used for this research. |
| Software Dependencies | No | The paper mentions 'Our model implementations were based on the MMDetection open source project (Chen et al., 2019)' and refers to specific techniques like '3x3 depth-wise separable convolution (Chollet, 2017)', but it does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | At each trial, we have trained the network for 15 epochs using Stochastic Gradient Descent (SGD) with momentum set to 0.9 and weight decay parameter set to 0.0001. The learning rate was set according to the linear scaling rule (Goyal et al., 2017); this rule states that the learning rate has to be proportional to the batch size, where each batch was set to include 2 input images, each at resolution of 1333 × 800 pixels. |