Mixed Supervised Object Detection by Transferring Mask Prior and Semantic Similarity

Authors: Yan Liu, Zhijie Zhang, Li Niu, Junjie Chen, Liqing Zhang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three benchmark datasets demonstrate the effectiveness of our method over existing methods.
Researcher Affiliation Academia Yan Liu , Zhijie Zhang , Li Niu , Junjie Chen, Liqing Zhang Mo E Key Lab of Artificial Intelligence Department of Computer Science and Engineering Shanghai Jiao Tong University {loseover, zzj506506, ustcnewly, chen.bys}@sjtu.edu.cn zhang-lq@cs.sjtu.edu.cn
Pseudocode Yes Algorithm 1 Iterative Training Strategy
Open Source Code Yes Codes are available at https://github.com/bcmi/Tra Ma S-Weak-Shot-Object-Detection.
Open Datasets Yes Following [45], we investigate COCO 2017 detection dataset [27] as our source dataset. Besides, we choose ILSVRC 2013 detection dataset [32] as another source dataset. We take Pascal VOC 2007 [10] as the target dataset
Dataset Splits Yes COCO 2017 is composed of an official train/validation split with 118287 and 5000 images with 80 categories covering the 20 categories in VOC. After that, we denote the resulting COCO 2017 as COCO-60 which consists of 21987 training images and 921 validation images. Both the training set and validation set are merged together as the source dataset. ... The 200-category dataset, including the 20 categories in VOC, is split into training (395909 images) and validation (20121 images) sets. ... Pascal VOC 2007 [10] as the target dataset, which is split into trainval (5011 images) and test (4952 images) sets.
Hardware Specification Yes All experiments are conducted on two 24GB TITAN RTX, with a batch size of 16 images.
Software Dependencies No The paper mentions using Faster RCNN, ResNet-50, and SGD optimizer, and cites PyTorch, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes The learning rate is initialized to 8 × 10−3 and reduced to 8 × 10−4. The weight decay and momentum are set to 1 × 10−4 and 0.9, respectively. The random seed is set to 222. ... we conduct four iterations (T = 4) following [45]. The threshold used to obtain candidate bounding boxes is 0.05. The threshold used to obtain pseudo bounding boxes of novel categories is 0.8. ... we use five image scales {480, 576, 688, 864, 1200} for both training and testing unless otherwise specified. ... α and γ are hyper-parameters set as 0.1 and 1.0 via cross-validation, respectively. ... β is a hyper-parameter set as 0.1 via cross-validation. ... In our experiments, we set both K and M to 8, so B = K × M = 64.