Skip to content
C3D is a modified version of BVLC tensorflow to support 3D ConvNets.
Branch: master
Clone or download
hx173149 Merge pull request #68 from frankgu/patch-1
Add a new version of C3D  that can use AWS Sagemaker
Latest commit b0b5389 Jul 9, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
C3D-tensorflow-1.0 Add files via upload Oct 18, 2017
list list/clear the train.list and the test.list files first Dec 29, 2016
LICENSE add LICENSE file Oct 16, 2017 Extensive tidy-up; Combined two video loading functions. Nov 28, 2016
crop_mean.npy add crop mean file Oct 19, 2017 convert the code to TF 1.2 version Oct 16, 2017 convert the code to TF 1.2 version Oct 16, 2017


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.


  1. Have installed the tensorflow >= 1.2 version
  2. You must have installed the following two python libs: a) tensorflow b) Pillow
  3. You must have downloaded the UCF101 (Action Recognition Data Set)
  4. Each single avi file is decoded with 5FPS (it's depend your decision) in a single directory.
    • you can use the ./list/ script to decode the ucf101 video files
    • run ./list/ .../UCF101 5
  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/ script to generate the {train,test}.list for the dataset
    • run ./list/ .../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


  1. python will train C3D model. The trained model will saved in models directory.
  2. python will test C3D model on a validation data set.
  3. cd ./C3D-tensorflow-1.0 &&python will get the random-clip accuracy on UCF101 test set with provided sports1m_finetuning_ucf101.model.
  4. C3D-tensorflow-1.0/ code is compatible with tensorflow 1.0+ , with a little bit different with the old repository
  5. 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 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.'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

Other verions


You can’t perform that action at this time.