C3D-tensorflow

This is a repository trying to implement C3D-caffe on tensorflow,useing models directly converted from original C3D-caffe.
Be aware that there are about 5% video-level accuracy margin on UCF101 split1 between our implement in tensorflow and the original C3D-caffe.

Requirements:

Have installed the tensorflow >= 1.2 version
You must have installed the following two python libs: a) tensorflow b) Pillow
You must have downloaded the UCF101 (Action Recognition Data Set)
Each single avi file is decoded with 5FPS (it's depend your decision) in a single directory.
- you can use the ./list/convert_video_to_images.sh script to decode the ucf101 video files
- run ./list/convert_video_to_images.sh .../UCF101 5
Generate {train,test}.list files in list directory. Each line corresponds to "image directory" and a class (zero-based). For example:
- you can use the ./list/convert_images_to_list.sh script to generate the {train,test}.list for the dataset
- run ./list/convert_images_to_list.sh .../dataset_images 4, this will generate test.list and train.list files by a factor 4 inside the root folder

database/ucf101/train/ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c01 0
database/ucf101/train/ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c02 0
database/ucf101/train/ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c03 0
database/ucf101/train/ApplyLipstick/v_ApplyLipstick_g01_c01 1
database/ucf101/train/ApplyLipstick/v_ApplyLipstick_g01_c02 1
database/ucf101/train/ApplyLipstick/v_ApplyLipstick_g01_c03 1
database/ucf101/train/Archery/v_Archery_g01_c01 2
database/ucf101/train/Archery/v_Archery_g01_c02 2
database/ucf101/train/Archery/v_Archery_g01_c03 2
database/ucf101/train/Archery/v_Archery_g01_c04 2
database/ucf101/train/BabyCrawling/v_BabyCrawling_g01_c01 3
database/ucf101/train/BabyCrawling/v_BabyCrawling_g01_c02 3
database/ucf101/train/BabyCrawling/v_BabyCrawling_g01_c03 3
database/ucf101/train/BabyCrawling/v_BabyCrawling_g01_c04 3
database/ucf101/train/BalanceBeam/v_BalanceBeam_g01_c01 4
database/ucf101/train/BalanceBeam/v_BalanceBeam_g01_c02 4
database/ucf101/train/BalanceBeam/v_BalanceBeam_g01_c03 4
database/ucf101/train/BalanceBeam/v_BalanceBeam_g01_c04 4
...

Usage

python train_c3d_ucf101.py will train C3D model. The trained model will saved in models directory.
python predict_c3d_ucf101.py will test C3D model on a validation data set.
cd ./C3D-tensorflow-1.0 &&python Random_clip_valid.py will get the random-clip accuracy on UCF101 test set with provided sports1m_finetuning_ucf101.model.
C3D-tensorflow-1.0/Random_clip_valid.py code is compatible with tensorflow 1.0+ , with a little bit different with the old repository
IMPORTANT NOTE: when you load the sports1m_finetuning_ucf101.model,you should use the tranpose operation like: pool5 = tf.transpose(pool5, perm=[0,1,4,2,3]),or in Random_clip_valid.py looks like:["transpose", [0, 1, 4, 2, 3]], but if you load conv3d_deepnetA_sport1m_iter_1900000_TF.model or c3d_ucf101_finetune_whole_iter_20000_TF.model,you don't need tranpose operation,just comment that line code.

Experiment result:

Note:
1.All report results are done specific on UCF101 split1 (train videos:9537,test videos:3783).
2.ALL the results are video-level accuracy,unless stated otherwise.
3.We follow the same way to extract clips from video as the C3D paper saying:'To extract C3D feature, a video is split into 16 frame long clips with a 8-frame overlap between two consecutive clips.These clips are passed to the C3D network to extract fc6 activations. These clip fc6 activations are averaged to form a 4096-dim video descriptor which is then followed by an L2-normalization'
C3D as feature extractor:

platform	feature extractor model	fc6+SVM	fc6+SVM+L2 norm
caffe	conv3d_deepnetA_sport1m_iter_1900000.caffemodel	81.99%	83.39%
tensorflow	conv3d_deepnetA_sport1m_iter_1900000_TF.model	79.38%	81.44%
tensorflow	c3d_ucf101_finetune_whole_iter_20000_TF.model	79.67%	81.33%
tensorflow	sports1m_finetuning_ucf101.model	82.73%	85.35%

finetune C3D network on UCF101 split1 with pre-trained model:

platform	pre-trained model	train-strategy	video-accuracy	clip-accuracy	random-clip
caffe	c3d_ucf101_finetune_whole_iter_20000.caffemodel	directly test	-	79.87%	-
tensorflow	c3d_ucf101_finetune_whole_iter_20000_TF.model	directly test	78.35%	72.77%	57.15%
tensorflow-A	conv3d_deepnetA_sport1m_iter_1900000_TF.caffemodel	whole finetuning	76.0%	71%	69.8%
tensorflow-B	sports1m_finetuning_ucf101.model	freeze conv,only finetune fc layers	79.93%	74.65%	76.6%

Note:
1.the tensorflow-A model corresponding to the original C3D model pre-trained on UCF101 provided by @ hx173149 .
2.the tensorflow-B model is just freeze the conv layers in tensorflow-A and finetuning four more epochs on fc layers with learning rate=1e-3.
3.the random-clip column means random choose one clip from each video in UCF101 test split 1 ,so the result are not so robust.But according to the Law of Large Numbers,we may assume this items is positive correlated to your video-level accuracy.
4.it's obvious that it if you do more finetuning work based on c3d_ucf101_finetune_whole_iter_20000_TF.model,and you may achieve better performance,i didn't do it because of time limit.
5.with no doubt that you can get better result by appropriately finetuning the network

Trained models:

Model	Description	Clouds	Download
C3D sports1M TF	C3D sports1M converted from caffe C3D	Dropbox	C3D sports1M
C3D UCF101 TF	C3D UCF101 trained model converted from caffe C3D	Dropbox	C3D UCF101
C3D UCF101 TF train	finetuning on UCF101 split1 use C3D sports1M model by @ hx173149	Dropbox	C3D UCF101 split1
split1 meanfile TF	UCF101 split1 meanfile converted from caffe C3D	Dropbox	UCF101 split1 meanfile
everything above	all four files above	baiduyun	baiduyun

References: