RQ1 What are the different error metrics present for functionality testing the depth estimation algorithms?
RQ2 Study of cost volume and Gaussian process on MVS
RQ3 Cross application of MVS approach
The original paper used only 4 error metrics ((i) L1, (ii) L1-rel, (iii) L1-inv, and (iv) sc-inv)to evaluate the performance of the model. We can evaluate the model with other metrics(squared distance loss; squared distance loss in log space; scale invariant loss by Eigen et al.; normalized loss;) used for performance measures.
About the cost volume, from what I read, we can incorporate cost volume at the decoder side and not on the input of the encoder. The attached image of the original paper and adding new cost volume at the decoder side improve the resolution of the depth image. (This itself is a paper written by 7 people).
And we have a cross-application of MVS (Extension to segmentation). Here we extract depth and feed it into the separate encoder and decoder architecture.
Another approach is the segnet, which helps to find the segmentation from images
Refined questions what I was thinking is
RQ1 What are the different error metrics present for functionality testing the depth estimation algorithms?
RQ2 Is it feasible to incorporate an adaptive thin volume on the decoder to improve the image resolution?
RQ3 Can we extend the temporal depth estimation using multi-view stereo to the segmentation task?
or study of gaussian with different kernels to multi-view stereo.
RQ1 What are the different error metrics present for functionality testing the depth estimation algorithms? RQ2 Study of cost volume and Gaussian process on MVS RQ3 Cross application of MVS approach
The original paper used only 4 error metrics ((i) L1, (ii) L1-rel, (iii) L1-inv, and (iv) sc-inv)to evaluate the performance of the model. We can evaluate the model with other metrics(squared distance loss; squared distance loss in log space; scale invariant loss by Eigen et al.; normalized loss;) used for performance measures.
About the cost volume, from what I read, we can incorporate cost volume at the decoder side and not on the input of the encoder. The attached image of the original paper and adding new cost volume at the decoder side improve the resolution of the depth image. (This itself is a paper written by 7 people).
And we have a cross-application of MVS (Extension to segmentation). Here we extract depth and feed it into the separate encoder and decoder architecture.
Another approach is the segnet, which helps to find the segmentation from images
Refined questions what I was thinking is
RQ1 What are the different error metrics present for functionality testing the depth estimation algorithms? RQ2 Is it feasible to incorporate an adaptive thin volume on the decoder to improve the image resolution? RQ3 Can we extend the temporal depth estimation using multi-view stereo to the segmentation task?
or study of gaussian with different kernels to multi-view stereo.
What is your feedback on the same?