Abstract: 自己记录一下每天Master项目的Log

2022

Jun

Upto 02

Progress

• Compute the proportion synthetic points, resemble real dataset metadata.
• create mapping function for the synthetic -> real data
• train squeezeSegV2 with synthetic data for 25 epochs
  • Tensorboard
• fine-tune Syn_Epoch25 with semanticKitti for 50Epoch
  • Tensorboard

Memo

• Train the model with transfer tech. does not work (Orange: real data tuning, Blue: synthetic data pretrain)

  

  

  

  

  By looking at the board, we can see that the model is not learning properly, as if it is well pretrained, the starting point for the tuning (orange) line should be closer to the blue one, as we assume two domains should share some features in higher dimension, like this:

  

  We are expecting the model to at least learn the intersection part of the knowledge from the synLiDAR, but in practice this is not the case, as shown by training progress.

• The width of the canvas set in the SqueezeSegv2 is 2048, which may be a bit too big for SynLiDAR. -> Going to change the size to 1024 and train again

• Going to use the pre-trained synLiDAR model to infer on the SemanticKitti directly, and test the performance:

  • A more detailed setting:

    train	FineTune?	Infer
    Synthetic	Synthetic	Real
    Synthetic	Real	Real
    Real	Real	Real

  • Going to evaluate above settings step by step.

06

Progress

• Train the RangeNet++ backboned with SqueezeSegv2 for Semantic Segmentation
  • real dataset
  • synthetic dataset
  • -> as step to row 1 and 3 for the transfer learning part

Schedule:

• Inference syn on real directly
• Inference real on real directly
• Inference real on syn directly
• Inference syn on syn directly

18

The biggest problem for choosing a good baseline model is that you do not know which trick they actually played to achieve a better scoring for the accuracy, and it is often the case that such technique is not applicable for other dataset, that is, they cannot be transplanted to other tasks at all..