RDSNet: A New Deep Architecture forReciprocal Object Detection and Instance Segmentation
Authors: Shaoru Wang, Yongchao Gong, Junliang Xing, Lichao Huang, Chang Huang, Weiming Hu12208-12215
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental analyses and comparisons on the COCO dataset demonstrate the effectiveness and efficiency of RDSNet. |
| Researcher Affiliation | Collaboration | Shaoru Wang,1,4 Yongchao Gong,2 Junliang Xing,1 Lichao Huang,2 Chang Huang,2 Weiming Hu1,3,4 1National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences 2Horizon Robotics Inc., 3CAS Center for Excellence in Brain Science and Intelligence Technology 4University of Chinese Academy of Sciences |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/wangsr126/RDSNet. |
| Open Datasets | Yes | We report the results on COCO dataset (Lin et al. 2014) |
| Dataset Splits | Yes | We train on train2017, and evaluate on val2017 and test-dev. |
| Hardware Specification | Yes | P means Titan XP or 1080Ti, and V means Tesla V100. |
| Software Dependencies | No | The paper states 'We implement RDSNet based on mmdetection (Chen et al. 2019b)' but does not provide specific version numbers for mmdetection or other software dependencies. |
| Experiment Setup | Yes | Dimensions of the instance and pixel representations are 32. We use different expanding ratios of bounding boxes for cropping masks during training and inference. During training, we use ground-truth bounding boxes and expand both the heights and the widths of them by 1.5 times with center point retaining. During inference, the expanding ratio is set to 1.2. All λs are set to 1. We train our models on 4 GPUs (2 images per GPU) and adopt 1 training strategy (Chen et al. 2019b) along with all other settings same as Retina Net, and then parameters in MBRM are trained individually for another 1k iterations. |