ViDT: An Efficient and Effective Fully Transformer-based Object Detector

Authors: Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo, Wonjae Kim, Ming-Hsuan Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive evaluation results on the Microsoft COCO benchmark dataset demonstrate that Vi DT obtains the best AP and latency trade-off among existing fully transformer-based object detectors, and achieves 49.2AP owing to its high scalability for large models.
Researcher Affiliation Collaboration 1NAVER AI Lab 2Google Research 3University of California at Merced 4Yonsei University
Pseudocode No The paper includes figures illustrating architectures but no formal pseudocode or algorithm blocks with numbered steps formatted as code.
Open Source Code Yes We release the code and trained models at https://github.com/naver-ai/vidt.
Open Datasets Yes We carry out object detection experiments on the Microsoft COCO 2017 benchmark dataset (Lin et al., 2014).
Dataset Splits Yes All the fully transformer-based object detectors are trained on 118K training images and tested on 5K validation images following the literature (Carion et al., 2020).
Hardware Specification Yes All the algorithms are implemented using Py Torch and executed using eight NVIDIA Tesla V100 GPUs.
Software Dependencies No The paper mentions PyTorch and AdamW but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We train Vi DT for 50 epochs using Adam W (Loshchilov & Hutter, 2019) with the same initial learning rate of 10 4 for its body, neck and head. The learning rate is decayed by cosine annealing with batch size of 16, weight decay of 1 10 4, and gradient clipping of 0.1.