search

programmingLearner / MATF-architecture-details

Architecture Details for CVPR 19 paper: Multi-Agent Tensor Fusion for Contextual Trajectory Prediction
MIT License
57 stars 10 forks source link

Did the model be trained with different model? #3

Open ghost opened 4 years ago

ghost commented 4 years ago

Hi I'm interesting in your paper and work. Did you train the model when encountering different data set? Or you train a general model that can get good result on all data set? Thanks

programmingLearner commented 4 years ago

Hi Johnny,

Thank you very much for your interest!

The model is trained in one of the datasets, and it is fine-tuned for each different dataset to report the final result.

Best, Tianyang

ghost commented 4 years ago

Hi Tianyan,

Thank you for the response.

So which dataset is used for training dataset? Also, what did you do when fine-tuning? I'm struggling with cross-dataset performance.

Johnny

programmingLearner commented 4 years ago

e.g. you could first train it on the SDD dataset, which is the largest, then fine-tuning on NGSIM / ETH-UCY dataset. During training, you might like to first train a single-agent LSTM, then multi-agent, and finally the whole MATF model in the paper, since we are all using residual learning, which will be easier if you are struggling with the performance. In our work, we also used the private MA highway dataset.