The results dont match with ones in the paper.

takumayagi / fpl

Code/data of the paper "Future Person Localization in First-Person Videos" (CVPR2018)

https://artilects.net/projects.html

MIT License

67 stars 16 forks source link

The results dont match with ones in the paper. #7

Open trungmanhhuynh opened 4 years ago

trungmanhhuynh commented 4 years ago

Hi @takumayagi,

I tried to run your codes by using exact given instructions for the proposed model only (cnn_ego_pose_scale) and got the results on First-Person Locomotion Dataset below:

If I take averages of all splits, it is clear that my numbers are not the same (much higher) as those you reported in your paper. Toward Away Across Average
109.03 75.56 93.10 77.26

Could you please point me to some directions that I can generate the same numbers that you reported?

Thank you so much,

sde123 commented 4 years ago

@trungmanhhuynh Hello, Did you occur the following problems when you run the command "python3 utils/run.py scripts/5fold.json run "?

"Written to experiments/5fold_191219_143815/191219_151658 Traceback (most recent call last): File "train_cv.py", line 55, in True, "scale" in args.model, args.ego_type) File "/home/dai/code/human_prediction_first_view/2/fpl/utils/dataset.py", line 91, in init ego_dict = data["egomotion_dict"][vid] KeyError: b'GOPR0234U20'" Looking forward to your reply!