HS-FPN: High Frequency and Spatial Perception FPN for Tiny Object Detection
Authors: Zican Shi, Jing Hu, Jie Ren, Hengkang Ye, Xuyang Yuan, Yan Ouyang, Jia He, Bo Ji, Junyu Guo
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments demonstrate that detectors based on HS-FPN exhibit competitive advantages over state-of-the-art models on the AI-TOD dataset for tiny object detection. ... We integrate HS-FPN into various detection models to replace FPN, and evaluate its performance on multiple tiny object detection datasets. Experiments indicate that HS-FPN significantly improves performance compared to FPN. ... Experiments are performed on two TOD datasets. The main experiments are conducted on the challenging AI-TOD (Wang et al. 2021) ... We analyze the effect of each proposed component of HS-FPN on AI-TOD test subset. Results are reported in Table 1. |
| Researcher Affiliation | Collaboration | 1Huazhong University of Science and Technology, Wuhan, China 2National Key Laboratory of Science and Technology on Multi-Spectral Information Processing, Wuhan, China 3Air Force Early Warning Academy, Wuhan, China 4Chinese People s Liberation Army 95841 troops, China EMAIL EMAIL |
| Pseudocode | No | The paper describes the HFP and SDP modules in detail with architectural diagrams (Figure 3, Figure 6) and textual descriptions, but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | No explicit statement about providing source code or a link to a code repository is found in the paper. The link provided in the 'Extended version' section is to an arXiv preprint (https://arxiv.org/abs/2412.10116), not a code repository. |
| Open Datasets | Yes | Experiments are performed on two TOD datasets. The main experiments are conducted on the challenging AI-TOD (Wang et al. 2021), which provides a total of 700,621 instances in 8 categories in 28,036 aerial images. ... Furthermore, we build a subset consisting of 10 categories named DOTAmini10 from DOTA (Xia et al. 2018) and test our method on it. |
| Dataset Splits | No | The paper mentions evaluating on the AI-TOD test subset and states that 'Models are trained on the train-val set and evaluated on the test set' (Table 3 caption), but it does not provide specific percentages, sample counts, or a detailed methodology for how these splits are created or defined beyond using existing dataset divisions. |
| Hardware Specification | Yes | We train and evaluate the detectors on two NVIDIA 3080 Ti GPUs (one image per GPU). |
| Software Dependencies | No | All of our experiments are implemented based on MMDetection (Chen et al. 2019) and Py Torch (Paszke et al. 2019). While these software packages are mentioned, specific version numbers for MMDetection and PyTorch are not provided, which is necessary for reproducibility. |
| Experiment Setup | Yes | All detectors are trained using the Stochastic Gradient Descent (SGD) optimizer for 12 epochs with a momentum of 0.9 and weight decay of 0.0001. The initial learning rate of two-stage models like Faster RCNN (Ren et al. 2015) and one-stage models like Retina Net (Lin et al. 2020) is set to 0.005 and 0.001 respectively, and decays at the 8th and 11th epochs. Besides, the basic anchor size is set to 2 for all anchor-based models. |