Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
DoGA: Enhancing Grounded Object Detection via Grouped Pre-Training with Attributes
Authors: Yang Liu, Feng Hou, Yunjie Peng, Gangjian Zhang, Yao Zhang, Dong Xie, Peng Wang, Yang Zhang, Jiang Tian, Zhongchao Shi, Jianping Fan, Zhiqiang He
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that training with synonymous attribute-based prompts allows Do GA to generalize multi-granular prompts and surpass previous state-of-the-art approaches, yielding 50.2 on the COCO and 38.0 on the LVIS benchmarks under the zero-shot setting. |
| Researcher Affiliation | Collaboration | 1 Institute of Computing Technology (ICT), Chinese Academy of Sciences, 2 University of Chinese Academy of Sciences, 3 Beihang University, 4AI Lab, Lenovo Research, 5Lenovo Ltd. |
| Pseudocode | Yes | Algorithm 1: Mixed Category Prompts for Plain Detection; Algorithm 2: Mixed caption Prompts for Phrase Grounding |
| Open Source Code | Yes | Code https://github.com/liuyang-ict/Do GA |
| Open Datasets | Yes | Extensive experiments demonstrate that training with synonymous attribute-based prompts allows Do GA to generalize multi-granular prompts and surpass previous state-of-the-art approaches, yielding 50.2 on the COCO and 38.0 on the LVIS benchmarks under the zero-shot setting. Based on the COCO dataset (Lin et al. 2014) that contains 80 known classes. LVIS (Gupta, Dollar, and Girshick 2019) datasets. Obj365 (Shao et al. 2019) and 1.3K classes in GQA (Hudson and Manning 2019). Flickr30k Ref COCO Ref COCO+ Ref COCOg. |
| Dataset Splits | Yes | COCO val, LVIS minival, Flickr30k Ref COCO Ref COCO+ Ref COCOg R@1 / R@5 / R@10 val test A test B val test |
| Hardware Specification | Yes | We use 24 V100 GPUs and batch size 24 to train with group operations and 16 V100 GPUs and batch size 48 without them. |
| Software Dependencies | No | The paper mentions software like 'spa Cy' and 'NLTK' and models like 'GPT-4', 'Swin Tiny', and 'BERT-base', but does not provide specific version numbers for any software libraries or dependencies used in the experimental setup. |
| Experiment Setup | Yes | We train Do GA-T with 12 training epochs and use Swin Tiny (Liu et al. 2021) and BERT-base (Devlin et al. 2019) as backbone. We use 24 V100 GPUs and batch size 24 to train with group operations and 16 V100 GPUs and batch size 48 without them. |