Abstract: Photometric stereo is a classic Shape-From-Shading method which reconstructs surfaces from lighting responses with multiple images. In this paper, we focus on robust example-based photometric stereo with noisy images. By applying a novel per-pixel noise detection approach before intensity vector comparison, the noise maps of every input image are acquired first. Then they are used to guide the comparison afterwards by ignoring noisy lighting intensity components. Our approach can handle images with lighting noises, either self-shadows or casted shadows. Furthermore, the approach relaxes the material requirement for the target object, only dichromatic reflection model is needed instead of Lambertian model. Also, lighting calibration is unnecessary, making the approach very easy to test and apply. Experiments on both synthetic data and real world images show that the approach gives more robust reconstruction results when the input images are noisy.