The article presents the LightCap model developed by Huawei for efficient image captioning, particularly focusing on mobile device deployment. It discusses the model architecture, training methodologies, and evaluations against state-of-the-art benchmarks. Key findings reveal that with a visual concept number set at K=20, the model maintains high performance while being optimized for mobile inference, processing images in approximately 188ms on a Huawei P40 with a Kirin 990 chip. This efficiency makes it suitable for practical applications in real-world scenarios, fulfilling the demand for speed and accuracy in mobile environments.
In our experiments, we found that the LightCap model achieved efficient inference on mobile devices, processing images in about 188ms on the Kirin 990 CPU.
The model showed promising results for real-world applications, combining efficiency with state-of-the-art performance in image captioning tasks, particularly when optimized for mobile usage.
Collection
[
|
...
]