FP-DETR: Detection Transformer Advanced by Fully Pre-training

Authors: Wen Wang, Yang Cao, Jing Zhang, Dacheng Tao

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the challenging COCO dataset demonstrate that our FP-DETR achieves competitive performance. Moreover, it enjoys better robustness to common corruptions and generalization to small-size datasets than state-of-the-art detection transformers.
Researcher Affiliation Collaboration 1University of Science and Technology of China 2Institute of Artificial Intelligence, Hefei Comprehensive National Science Center 3The University of Sydney, 4JD Explore Academy, China
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Code will be made publicly available at https://github.com/encounter1997/FP-DETR.
Open Datasets Yes Datasets. Following the common practice, our detector is pre-trained on Image Net (Deng et al., 2009) and fine-tuned on COCO 2017 (Lin et al., 2014) train set. ... Besides, we evaluated the model s generalization ability by fine-tuning on the small-size dataset, i.e., Cityscapes dataset (Cordts et al., 2016).
Dataset Splits No While the paper mentions fine-tuning on the 'COCO 2017 train set' and reporting 'Evaluation results on the val set of COCO 2017', it does not explicitly provide specific percentages, sample counts, or a detailed methodology for these dataset splits within the paper text itself.
Hardware Specification Yes All experiments are implemented on the NVIDIA A100 GPU.
Software Dependencies No The paper mentions optimizers like 'Adam W' but does not specify software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8').
Experiment Setup Yes By default, our FP-DETR is pre-trained on Image Net-1k (Deng et al., 2009) for 300 epochs with Adam W (Loshchilov & Hutter, 2018) optimizer and cosine learning rate scheduler. Training strategies in Dei T (Touvron et al., 2021a) are adopted, and the image size is set as 224 224. We use a batch size of 1,024 for training, and the initial learning rate is set as 5 10 4. After pre-training, models are fine-tuned for 50 epochs with Adam W optimizer on the downstream tasks. The learning rate is initialized as 1 10 4 and decreased by a factor of 0.1 at the 40th epoch. ... Besides, we set both the number of sampling points and the feature levels in multi-scale deformable attention as 4, and the number of object query embeddings as 300. Models are fine-tuned with a batch size of 32.