One-Shot Object Detection with Co-Attention and Co-Excitation

Authors: Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, Tyng-Luh Liu

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train and evaluate our model on VOC and COCO benchmark datasets. For VOC, our model is trained on the union set of VOC 2007 train&val sets and VOC 2012 train&val sets, and is evaluated on VOC 2007 test set. ... Table 1 shows that our model using reduced Image Net pre-trained backbone ( Ours (725) ) still achieves better performance on both seen and unseen classes than the baseline methods.
Researcher Affiliation Collaboration 1National Tsing Hua University, 2Academia Sinica, Taiwan, 3Aeolus Robotics, 4Taiwan AI Labs
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Codes are available at https://github.com/timy90022/One-Shot-Object-Detection.
Open Datasets Yes Following the previous work [22, 24], we train and evaluate our model on VOC and COCO benchmark datasets.
Dataset Splits Yes For VOC, our model is trained on the union set of VOC 2007 train&val sets and VOC 2012 train&val sets, and is evaluated on VOC 2007 test set. For COCO, our model is trained on COCO train 2017 set and evaluated on COCO val 2017 set.
Hardware Specification Yes We train our models using SGD optimizer with momentum 0.9 for ten epochs, with batch size 128 on eight NVIDIA V100 GPUs in parallel.
Software Dependencies No The paper mentions using Mask-RCNN and provides a link to an implementation, but it does not specify software dependencies with version numbers for its own method (e.g., Python, PyTorch/TensorFlow versions, or other libraries with versions).
Experiment Setup Yes We train our models using SGD optimizer with momentum 0.9 for ten epochs, with batch size 128 on eight NVIDIA V100 GPUs in parallel. We use a learning rate starting with 0.01, and then decay it by a ratio 0.1 for every four epochs. We use λ = 3 in (7) for the margin-based ranking loss.