We introduce a camera simulator to synthesize raw sensor data under different camera settings, including exposure time, ISO, and aperture. The simulator consists of three components: an exposure module relying on the principle of modern lens designs, a noise module utilizing deep convolutional denoising networks and noise level functions, and an aperture model using a new adaptive attention module. Through the proposed pipeline, we can correct the luminance level, adapt the noise, and synthesize the defocus blur. We collected a dataset of more than ten thousand raw images of 450 diverse scenes with different camera settings using two cameras. The dataset can not only facilitate the training of the simulator but also benefit other tasks such as training image descriptors. Quantit...[ Read more ]

We introduce a camera simulator to synthesize raw sensor data under different camera settings, including exposure time, ISO, and aperture. The simulator consists of three components: an exposure module relying on the principle of modern lens designs, a noise module utilizing deep convolutional denoising networks and noise level functions, and an aperture model using a new adaptive attention module. Through the proposed pipeline, we can correct the luminance level, adapt the noise, and synthesize the defocus blur. We collected a dataset of more than ten thousand raw images of 450 diverse scenes with different camera settings using two cameras. The dataset can not only facilitate the training of the simulator but also benefit other tasks such as training image descriptors. Quantitative comparisons and qualitative results demonstrate that our approach outperforms relevant baselines in raw data simulation. Furthermore, our camera simulator can enable multiple applications, including large-aperture enhancement, HDR, and training auto exposure mode. Our work represents the first effort to fully simulate a camera sensor's behavior, leveraging both the power of conventional raw sensor characteristics and the potential of data-driven deep learning.