Thank you for the amazing paper and the elaborate code. I am trying to replicate the results of your paper on the egogesture dataset. However, I am unable to get the accuracy even near to what you have achieved. Could you please help me figure out what I am doing wrong.
The parameters that I have used for training the classifier in the run-egogesture.sh are as follows:
Hello @ahmetgunduz, @parkjh688, @okankop
Thank you for the amazing paper and the elaborate code. I am trying to replicate the results of your paper on the egogesture dataset. However, I am unable to get the accuracy even near to what you have achieved. Could you please help me figure out what I am doing wrong.
The parameters that I have used for training the classifier in the run-egogesture.sh are as follows:
python3 main.py --root_path ~/dataset/Real-time-GesRec-master \ --video_path datasets/egoges/images \ --annotation_path annotation_EgoGesture/egogestureall_but_None.json \ --result_path results_final/ego \ --pretrain_path pretrained_models/jester_resnext_101_RGB_32.pth \ --dataset egogesture \ --n_classes 27 \ --n_finetune_classes 83 \ --model resnext \ --width_mult 1.0 \ --model_depth 101 \ --resnet_shortcut B \ --resnext_cardinality 32 \ --train_crop random \ --learning_rate 0.01 \ --sample_duration 32 \ --modality RGB \ --pretrain_modality RGB \ --downsample 1 \ --batch_size 24 \ --n_threads 16 \ --checkpoint 1 \ --n_val_samples 1 \ --n_epochs 30 \ --ft_portion complete \ --test \ --lr_steps 10 25 \ --test_subset test \ --train_validate \
However the best precision and recall values for 30th epochs are as follows: