Indiscernible Object Counting in Underwater Scenes�Guolei Sun, Zhaochong An, Yun Liu, Ce Liu,
Christos Sakaridis, Deng-Ping Fan*, Luc Van Gool
1
,
3
�
Introduction:
- Indiscernible object recognition
- camouflaged object detection, camouflaged instance segmentation, video camouflaged object detection
- Existing object counting tasks: Generic Object Counting and Dense Object Counting
- No previous research has focused on counting indiscernible objects
IOCFormer
- Mainstream methods for object counting
- Counting-by-density or counting-by-regression
- IOCFormer
Experiments
- Evaluated 14 existing popular counting methods
- Our method achieves SOTA performance
Summary/Conclusion
- Provide a rigorous study of the new task of indiscernible object counting
- Present the high-quality IOCfish5K dataset in underwater scenes
- Provide extensive benchmarking experiments and a new method
Code
https://github.com/GuoleiSun/Indiscernible-Object-Counting
Contribution:
- Propose the new task of indiscernible object counting (IOC).
- Contribute a large-scale dataset, named IOCfish5K, containing 5,637 images and 659,024 accurate point labels.
- Benchmark 14 classical and high-performing approaches for object counting on IOCfish5K
- Propose a novel approach, IOCFormer, which achieves state-of-the-art results in IOC
- Integrate density-based and regression-based methods in a unified framework
- Propose a density-based transformer encoder to exploit the object density information (the density branch) in the regression branch