We compare ClearSplatting-2.0 against Dex-NeRF, Res-NeRF, 3DGS, and ClearSplatting by computing the MAE and RMSE (figure 10).
Here i ∈ [0,…,N] is the frame number, r is the pixel location, and Ωr is the set of all pixel locations across frames. Dˆ(r) is the inferred depth in meters, D(r) is the GT depth in meters. We crop each image before evaluation to focus on the transparent object and not bias the results with the background.
The objective results based on the metrics, RMSE and MAE, can be found in the tables below.
From Figure 9, 10, and 11, we see that Clear-Splatting outperforms NeRF-based baselines as well as other 3DGS-based baselines. The results suggest that Clear-Splatting improves on the NeRF-based approaches with a 67.09% lower RMSE and an 87.80% lower MAE in depth estimation. ClearSplatting-2.0 beats Clear-Splatting by upto 33% lower RMSE and by upto 32% lower MSE in depth estimation.
From Figure 9, we see that the depth maps obtained using NeRF approaches do not have crisp boundaries in contrast to those obtained using 3DGS, ClearSplatting, and ClearSplatting-2.0. Furthermore, the depth maps obtained from baseline 3DGS has holes which can potentially lead to incorrect gripper pose estimation. ClearSplatting-2.0 is able to considerably close the holes obtained in the baselines and consequently lead to better gripper pose estimation.