README.md

To Reproduce Our Action Localization Results

1.Download Features and Annotations

Download the mammalnet_features for localization

wget https://mammalnet.s3.amazonaws.com/full_video_features/mammalnet_feature.tar.gz

The annotation file can be downloaded from here:

wget https://mammalnet.s3.amazonaws.com/annotation.tar.gz

Details: The file includes I3D features and action annotations in json format (similar to ActivityNet annotation format). To produce the features for our MammalNet videos, we firstly format the all videos to 25 FPS, then finetune a two-stream I3D model, that is originally pretrained on ImageNet and Kinetics 400, on our dataset, and finally extract the RGB and optical flow features for each video. We concatenate these two features together as the model input. Feature extraction can refer to mmaction2 and I3D Feature Extraction.

• The feature folder structure is

Feature folder
└───mammalnet_feature/
│    └───RGB_feature/
│    │	 └───SMN6WFVy-Ys.npy
│    │	 └───Di4eEBZjkA4.npy   
│    │	 └───...
│    └───Flow_feature/
│    │	 └───SMN6WFVy-Ys.npy
│    │	 └───Di4eEBZjkA4.npy   
│    │	 └───...
│    └───Concatenate_feature/
│    │	 └───SMN6WFVy-Ys.npy
│    │	 └───Di4eEBZjkA4.npy   
│    │	 └───...

2.Model Code

• We trained all detection models using their officially released code: ActionFormer, TAGS, and CoLA.
• You only need to convert our annotation json and set the feature path based on the official folder and annotation file configuration.