BC04 Multi-Sphere Results (3DoF Demo Data)
Created on July 20|Last edited on July 20
Comment
For 4DoF results see https://wandb.ai/mooey5775/mixed_media/reports/BC04-Multi-Sphere-Results-4DoF-Demo-Data---VmlldzoyMzA3Njg3
Overall CommentsResultsExample GIFsDirect Vector MSEToolFlowNet (No SVD), 3DoFToolFlowNet, 3DoF, Observed
Overall Comments
See the 4DoF wandb for details on what the keywords mean. Recall that eval/info_done_final is what we normally report in papers.
Results
Direct Vector (MSE)
5
ToolFlowNet (no SVD)
5
ToolFlowNet (SVD)
5
Run set
15
Example GIFs
Direct Vector MSE
This seems way worse than the single-sphere case, relatively speaking? Maybe that's good news? :)
/Users/danielseita/tmp/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_ee_3DoF_ar_8_hor_100_rawPCL_scaleTarg/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_ee_3DoF_ar_8_hor_100_rawPCL_scaleTarg_2022_07_18_17_33_36_0001/video
At epoch 900. This got 8/25 success.
ToolFlowNet (No SVD), 3DoF
data/local/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_avg_ee_3DoF_ar_8_hor_100_rawPCL_scaleTarg/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_avg_ee_3DoF_ar_8_hor_100_rawPCL_scaleTarg_2022_07_18_18_00_07_0001
At 125 epochs, with getting 10/25 success. (Also, 8/25 of these show it getting the wrong ball, usually when the red ball moved near the ladle when the ladle was inserted into water). I picked 125 epochs as it was badly overfitting for some reason.
ToolFlowNet, 3DoF, Observed
From epoch 975:
BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_svd_pointwise_flow_3DoF_ar_8_hor_100_scalePCL_noScaleTarg/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_svd_pointwise_flow_3DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_07_18_17_31_13_0001/video
At epoch 975 it was actually doing pretty well. There are 12/25 successes and 7/25 cases when it gets just the distractor (wrong). For those 7 cases, 3 are in the bottom row. It seems like the distractor moved towards the ladle?
Here's another one which got 15/25 successes also at epoch 975:
/Users/danielseita/tmp/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_svd_pointwise_flow_3DoF_ar_8_hor_100_scalePCL_noScaleTarg/BC04_MMMultiSphere_v02_ntrain_0100_PCL_PNet2_svd_pointwise_flow_3DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_07_18_17_31_13_0002/video
Add a comment