Recently, progress in 3D reconstruction has advanced significantly, mainly driven by methods based on Gaussian Splatting and Neural Radiance Fields (NeRF). However, many such methods assume ideal capture conditions, where cameras can acquire input images from any position and environmental factors do not degrade their quality. Yet, in many real-world scenarios, these assumptions break down due to various limitations, such as environmental constraints or unconventional imaging surfaces. We focus on enhancing 3D reconstruction methods to operate effectively under such adverse capture conditions where conventional approaches fail, addressing three challenging scenarios:
1) underwater imaging, where range of motion is often restricted so cameras cannot acquire input images from arbitrary positions,
2) sonar imaging, where measurements are incomplete and corrupted by noise, and
3) non-line-of-sight (NLOS) imaging using corneal reflections, where the reflections are corrupted by the texture of the iris. We each of these challenging scenarios in detail, highlighting the specific limitations they present and how to address them using techniques spanning computational imaging, machine learning, and computer graphics.