This is a consecutive frame target detection project by adding convlstm to SSD.
By adding conv LSTM to SSD, the highest accuracy is improved by nearly 10%. We have carried out on the radar data set provided by 25 customers. There are three targets in total. The network improves target 2 and target 3 greatly, target 2 increases by 5%, target 3 increases by 10%, and target 1 only improves by less than 1%.
You can use the split_vidseq.py in ./data/ to make your own dataset. Since my data is confidential, it is not convenient to release it.
SSD constructure:
SSD+ConvLSTM constructure:
