Hello,
Many thanks for sharing your great work.
I noticed that in the code the input image is resized to 112 but in the paper, it was reported 320*240.
Can you please elaborate this issue?
Thanks
Mohammad

Hi @GibranBenitez

Firstly thank you for publishing the dataset and code.

I ran the isolated test successfully with your dataset and pretrained weights, but I got a different confusion matrix from the paper. So I am wondering if I missed sth, or just you published different splits?

Moreover, when I tried to load resnet50 with pretrained weights, the shapes do not match. Please tell me how to fix it. Thx.

When I was building dataset from video files (smaller download), I found that the number of frames in some files doesn't match the number of frames reported in the annotations or IPN_Hand/frames/<vid_name>_{:06d}.jpg.

for example: 1CM1_3_R_#226.avi

ffprobe -v error -select_streams v:0 -count_frames -show_entries stream=nb_read_frames -print_format csv 1CM1_3_R_#226.avi

stream,4786

However, final annotation ends with frame 4795

# Annot_List.txt
video,label,id,t_start,t_end,frames
...
1CM1_3_R_#226,D0X,1,4769,4795,27
...

and the folder of image frames from video also contains 4795 frames.

What method did you use to split the .avi videos into individual frames?

hello, i can not find egogesture.py in the folder datasets. should i delete the code "from datasets.egogesture import EgoGesture" ?

Hello,
Many thanks for sharing the code.
Can you please let me know how I can test the model on a custom dataset?
Can you please also clarify the difference between online and offline testing? From what I noticed, for both cases, we need annotations.
Thank you

Thanks for your great work.I am trying to reproduce the results of ResNeXt101 on IPN-hand, However, I am unable to do so and I would appreciate some help.
i followed the training script run_clf_ipn_trainRex-js32b32.sh under the folder test/,here are my python args

python main.py
--root_path . --video_path datasets/HandGestures/IPN_dataset --annotation_path annotation_ipnGesture/ipnall_but_None.json --result_path results_ipn --pretrain_path report_ipn/ResNeXt101/shared_models_v1/models/jester_resnext_101_RGB_32.pth --pretrain_dataset jester --dataset ipn --sample_duration 32 --learning_rate 0.01 --model resnext --model_depth 101 --resnet_shortcut B --batch_size 384 --n_classes 13 --n_finetune_classes 13 --n_threads 16 --checkpoint 1 --modality RGB --train_crop random --n_val_samples 1 --test_subset test --n_epochs 100 --store_name ipnClf_jes32r_b32

i trained for 100 epoch,my best val acc is 65%,In your paper it is 83%, and I was able to achieve around 83% validation accuracy using the pre-trained weights of ResNeXt101. However, I am unsure if I made any mistakes in my settings or if there are any specific training strategies that could improve the accuracy further.
opts_ipnClf_jes32r_b32_resnext-101.txt

Hii @GibranBenitez ,
I am facing trouble to draw inference from the model . Can you please share the steps how I can do the inferencing, that would be very helpful !!
You can also share me some reference so that I can try , any kind of help is appreciated !!

@GibranBenitez
While doing online test of the dataset I get the following error

It is caused due to the following line in online_test.py
elif opt.dataset == 'ipn':
file_set = os.path.join(opt.video_path, 'ValList.txt')
test_paths = []
buf = 0
with open(file_set,'rb') as f:
for line in f:
vid_name = line.decode().split('\t')[0]
test_paths.append(os.path.join(opt.video_path, 'frames', vid_name))

However I could not find this file anywhere in your drive or the repo. Could you please provide this file. Is it the same as the vallistall.txt file in the annotation_ipn directory