Objective | Techniques | Reference | Year | Limitations/research gap |
---|---|---|---|---|
3D mesh compression | DCT based source image compression | [151] | 2022 | Direct compression on a 3D model has not been performed |
Training neural networks with signed distance function and network weights as a field | [152] | 2022 | The method is not lossless as mapping weights back to coordinates is not reversible | |
3D mesh generation | Recurrent neural network (RNN) | [153] | 2022 | Supervised learning and requires training labels for target normal. There is still a gap in unsupervised polygon normal regression |
Deep normal filtering neural network | [149] | 2020 | ||
3D mesh encryption | Encryption using chaotic behavior in edge computing devices | [154] | 2023 | Even though vertices are encrypted the edges remain connected and limits model security |
Using correlation between two sets of vertices to recover encrypted data | [155] | 2022 | ||
3D mesh smoothing | RNN | [153] | 2022 | These methods either correct the vertices or correct the face normal. These methods do not output a final mesh with vertices and labelled edges |
Deep normal filtering neural network | [149] | 2020 | ||
Medical scan segmentation | U-Net | [63] | 2015 | Supervised learning and requires training labels for target masks. There is a huge gap in unsupervised DL based segmentation |
Attention U-Net | [156] | 2022 | ||
Real time 3D imaging based prediction and diagnosis | Spatio-temporal long short term memory | [157] | 2022 | Real time 3D model of a moving heart is not constructed in this paper. Faster and efficient way of 3D organ modeling needs to be studied as normal heart rate is 60–100 bpm. So each model should not be modeled fast |