Consultation on reproducing results

[zxzxzxzxzxzxzxzxzxzxzx]

instructions：nohup python -u main_linux.py --config ./configs/30_frames_15_stride/deepconvlstm_long.yaml --seed 1 --eval_type split > exp1.log 2>&1 &

main.py：

[mariusbock]

The results need to be postprocessed. Without applying the majority filter (see postprocessing/majority_fil.py) the mAP scores of the inertial models will be near zero due to frequently happening one-off mispredictions (ultimately splitting segments). Hope this helps!

[zxzxzxzxzxzxzxzxzxzxzx]

Okay, thank you very much for your answer and your work support. I will now debug majority_fil. py

[Peipi98]

Hello, I've checked the instructions, but if oracle, camera and inertial are clear to me, the combined approach in early fusion isn't. If I've correctly understood, the early fusion approaches has to be treated as well as the vision based, by using score_thresh.py: am I right or did you use another approach?

[mariusbock]

Yes you are correct. The early-fusion experiments were only performed with the Temporal Action Localization models (Tridet and ActionFormer). Hence, we have to apply score thresholding like we do in the vision-based experiments

[Peipi98]

thanks!