SpeedDETR: Speed-aware Transformers for End-to-end Object Detection

Authors: Peiyan Dong, Zhenglun Kong, Xin Meng, Peng Zhang, Hao Tang, Yanzhi Wang, Chih-Hsien Chou

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the MS COCO dataset show Speed DETR outperforms current DETR-based methods by 1.5% 9.2% AP with 1.09 3.6 speedup on Tesla V100.
Researcher Affiliation Collaboration 1Northeastern University, Boston, MA, U.S.A 2Futurewei Technologies, Santa Clara, CA, U.S.A 3Peking University, Beijing, China 4Tsinghua University, Beijing, China 5CVL, ETH, Switzerland.
Pseudocode Yes Algorithm 1 Speed-aware Model Slimming.
Open Source Code Yes Codes release Speed DETR.
Open Datasets Yes Experiments on the MS COCO dataset show Speed DETR outperforms current DETR-based methods by 1.5% 9.2% AP with 1.09 3.6 speedup on Tesla V100. and The HOD backbone is pre-trained on Image Net (Deng et al., 2009) with the same setting as (Cao et al., 2021).
Dataset Splits No The paper mentions training epochs and optimizer settings but does not explicitly provide details about validation dataset splits or how validation was performed.
Hardware Specification Yes Experiments on the MS COCO dataset show Speed DETR outperforms current DETR-based methods by 1.5% 9.2% AP with 1.09 3.6 speedup on Tesla V100. Even acceptable speed inference can be achieved on edge GPUs, i.e., 4 FPS for NVIDIA JETSON TX2 (1.4 4 faster than other counterparts), 1 FPS for NVIDIA NANO (1.5 6.7 faster). and We test four types of hardware devices, and their properties are listed in Table 1.
Software Dependencies No The paper mentions 'Py Torch Profiler' but does not provide specific version numbers for software dependencies like PyTorch itself or other libraries used in implementation.
Experiment Setup Yes We use the Adam W optimizer with a batch size of 32, an initial learning rate of 1e 4, and a weight decay of 0.05. The learning rate is stepped down by a factor of 0.1 at the 67% and 89% of training epochs.