CNN Regression for localisation

CNN architectures:

Full CNN training

• architecture as used for previous experiments
• replication of basic architecture (before motor inputs) in [1]
• replication of full architecture in [1]

All the above architectures achieve around 0.0002 MSE.

CNN codes

Achieves 0.002 MSE which is worse than when retraining the whole network. Possibly applying Global Average Pooling could help.

Conclusions

• drop out reduces performance a lot
• very complex archivectures on CNN codes only improved training MSE

Comments about data

05_3: no red gear
10_44: red gear on a side

Future work

• consider lower number of conv layers
• Global Average Pooling for visualising activation maps

Common issues

Faced this issue: tensorflow/tensorflow#6968 This helped: export LD_PRELOAD="/usr/lib/libtcmalloc_minimal.so.4"

References

[1] Levine, S., Pastor, P., Krizhevsky, A., Ibarz, J. and Quillen, D., 2016. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International Journal of Robotics Research, p.0278364917710318.