flowerfan / sf-net Goto Github PK

Hello, thanks for your interesting work SF-Net!
I have some detail question about the feature of the dataset GTEA and BEOID:
I notice that the feature dimension of each video is T*2048. From this paper, I get that this feature contains two parts (flow and RGB features). Can you tell me which part of the feature is the optical flow feature and which part is the RGB feature? Thanks for your help!

Hi Fang,

I am trying to extract I3D features from raw videos according to the description in Sec. 4.2. According to the sentence "The inputs to the I3D models are stacks of 16 frames", the feature size of one video should be [#frame / 16, 2048]. However, I got a different result. For example, “video_validation_0000001.mp4” in the THUMOS14 dataset contains 7845 frames. The feature size of it should be [491, 2048]. However, its feature downloaded from here is [408, 2048].

Are there other preprocessing steps? Or could you please share the feature extraction code?

Thanks.

Best,

Changjian

Nice work and thank you for sharing your code!
Could you please tell me about your computing resources on this task? E.g., how many and what type of GPU did you use and how long is the training time? Thank you!

Hello, thanks for sharing the implementation of this excellent work.

I have a question about the performance on BEOID.

When I run your code for the 2000 iterations as the paper specifies, the model scored a score of 11 % at the threshold of 0.7.

However, the reported score in your paper is much lower, which is 3.5 % of [email protected].

Could you please explain why the gap exists and the model is underestimated in the paper?

Thanks!

Due to some reasons, I cannot download the extracted features from google driven, especially for thumos, so could you please provide the baidu disk link for extracted features? Thanks very much!

hi, thank you for publishing this paper.

As I read this paper, I have a question.

1. Can I know the model used in Table 1 and the model used to measure Full in Table 3 ?

2. Can I know what hit means in Table 1 ?

Thank you for answering

Hi,

I was running the code without change on GTEA and I got the following output.

> python main.py --dataset-name=GTEA

# .....
Iteration: 7997, Loss: 5.519 
Iteration: 7998, Loss: 5.513 
Iteration: 7999, Loss: 5.524 
Iteration: 8000, Loss: 5.525 
model_name: sfnet
[INIT] Loaded annotations from validation subset.
	Number of ground truth instances: 130
	Number of predictions: 72
	Fixed distance to center ratio: [0.25 0.5  0.75 1.  ]
Warning: No predictions of label 'close' were provdied.
Warning: No predictions of label 'put' were provdied.
Warning: No predictions of label 'stir' were provdied.
	 mAP: 0.00102,0.00463,0.01775,0.03984
	Average-mAP: 0.01581063337015718
[INIT] Loaded annotations from validation subset.
	Number of ground truth instances: 130
	Number of predictions: 43
	Fixed threshold for tiou score: [0.1 0.2 0.3 0.4 0.5 0.6 0.7]
Warning: No predictions of label 'close' were provdied.
Warning: No predictions of label 'put' were provdied.
Warning: No predictions of label 'stir' were provdied.
	 mAP: 0.07106,0.04523,0.03729,0.02236,0.01005,0.00079,0.0
	Average-mAP: 0.026684358725175054
propsoal AP: 0.27856,0.25295,0.1789,0.06556,0.03295,0.01172,0.00506
99.65986410776775
All act frames 397, predict all frames : 72, right frames: 12,  AP: 0.16667, Recall: 0.03023

I guess there should be something wrong with my setup or the way I run the code.
For BEOID I was able to reproduce the results (~29% mAP).
I also tried it multiple times and always Average-mAP is around 0.02.

What do you think?

Dear author,

thanks for the great work, and when i set up all the env to test the code, i encounter with some issues, mostly happened in "video_dataset.py":

data = pd.read_csv(os.path.join(annotation_dire, filename), names=[
  File "/home/sc/anaconda3/envs/data/lib/python3.9/site-packages/pandas/util/_decorators.py", line 311, in wrapper

pandas.errors.ParserError: Too many columns specified: expected 3 and found 1

Do you have any idea about solving this?

Thanks a lot for the support.
sc