DTG-SSOD: Dense Teacher Guidance for Semi-Supervised Object Detection
Authors: Gang Li, Xiang Li, Yujie Wang, Wu Yichao, Ding Liang, Shanshan Zhang
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on COCO benchmark under Partially Labeled Data and Fully Labeled Data settings. |
| Researcher Affiliation | Collaboration | 1Nanjing University of Science and Technology 2Sense Time Research 3Nankai University |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code will be released at: https://github.com/ligang-cs/DTG-SSOD. |
| Open Datasets | Yes | We benchmark our proposed method on the challenging dataset, MS COCO [34]. |
| Dataset Splits | Yes | The val2017 set with 5k images is used as the validation set. We describe two training settings as follows: Partially Labeled Data. The train2017 set, consisting of 118k labeled images, is used as the training dataset, from which, we randomly sample 1%,2%,5%, and 10% images as labeled data, and set the remaining unsampled images as unlabeled data. Following the practice of previous methods [28, 9, 13], for each labeling ratio, 5 different folds are provided and the final result is the average of these 5 folds. |
| Hardware Specification | Yes | The model is trained for 180k iterations on 8 V100 GPUs with an initial learning rate of 0.01 |
| Software Dependencies | No | The paper mentions using Faster RCNN, FPN, ResNet50, and SGD but does not specify software dependencies with version numbers (e.g., PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | The model is trained for 180k iterations on 8 V100 GPUs with an initial learning rate of 0.01, which is then divided by 10 at 120k iteration and again at 160k iteration. Mini-batch size per GPU is 5, with 1 labeled image and 4 unlabeled images. The loss weight of unlabeled images α is set to 4.0. ... We set τ to 0.9 in NMS of the RPN stage, and 0.45 in NMS of the R-CNN stage, empirically. |