search

wei-tim / YOWO

You Only Watch Once: A Unified CNN Architecture for Real-Time Spatiotemporal Action Localization
Other
847 stars 158 forks source link

Train on jhmdb21 #25

Closed yzfly closed 4 years ago

yzfly commented 4 years ago

Hello, thank you for your excellent work.

I have some problems when training YOWO on jhmdb21 dataset:

  1. How to convert jhmdb21 Ground Truth (joint_positions) .mat file to .txt just like ucf24/labels/ ?

  2. How to get video_mAP? I have seen the video_mAP.py source code and find that I don't have "testlist_video.txt" file. How to generate such a file? or can you share yours? testlist = os.path.join(base_path, 'testlist_video.txt')

  3. In train.py, is locolization or localization? which one is right? locolization_recall = 1.0 * total_detected / (total + eps) print("Classification accuracy: %.3f" % classification_accuracy) print("Locolization recall: %.3f" % locolization_recall)

yzfly commented 4 years ago

I found related issue: issue18 I downloaded all the files you shared on dropbox. And did't find training dataset annotations. just as image

okankop commented 4 years ago

@yzfly The annotations in groundtruths_ucf.zip is only used at calculating the Frame_mAP scores. This is why there is no training annotations there. Training annotations can be downloaded from the dataset sources.

okankop commented 4 years ago

@yzfly Please use the following steps to use the codes in the repo. These steps have recently been validated by another user of the repo! They will also answer your questions above.

FOR UCF DATASET:

For Training

  1. Download ucf dataset, unzip file. Rename the datafolder as ‘ucf24’.
  2. Create a new folder by the name of ‘datasets’. Put the above renamed ‘ucf24’ folder inside the datasets folder.
  3. Download code from GitHub page. The code folder is called YOWO
  4. Open cfg/ucf24.data’ in the YOWO folder. Update address of the base, training and validation dataset
  5. Download ‘trainlist.txt’ and ‘testlist.txt’ files and put it inside the ‘ucf24’ folder created in step-1.
  6. Create a folder called ‘weights’ in the YOWO folder.
  7. Download pre-trained yolo weights, details given the GitHub page, and put it in the folder created in step-6.
  8. Download ‘resnext-101-kinetics.pth’ from the GitHub page and put in the ‘weights’ folder created in step -6.

For Validation

  1. Go to YOWO/evaluations/Object-Detection-Metrics
  2. Unzip the file ‘groundtruths_ucf.zip’
  3. Go back to the YOWO folder. Create a folder called ‘ucf_detections/detections_0’. Keep creating multiple subdirectories ‘detections_n’ for n = 0,1,2,3….. for further tasks.
  4. Run validation using the command

Run both training and validating model using the script: $ sh run_ucf101-24.sh

  1. Change the ‘detection_n’ portion in the script file for multiple validation tasks. python ./evaluation/Object-Detection-Metrics/pascalvoc.py --gtfolder groundtruths_ucf --detfolder ../../ucf_detections/detections_0
yzfly commented 4 years ago

@yzfly Please use the following steps to use the codes in the repo. These steps have recently been validated by another user of the repo! They will also answer your questions above.

FOR UCF DATASET:

For Training

  1. Download ucf dataset, unzip file. Rename the datafolder as ‘ucf24’.
  2. Create a new folder by the name of ‘datasets’. Put the above renamed ‘ucf24’ folder inside the datasets folder.
  3. Download code from GitHub page. The code folder is called YOWO
  4. Open cfg/ucf24.data’ in the YOWO folder. Update address of the base, training and validation dataset
  5. Download ‘trainlist.txt’ and ‘testlist.txt’ files and put it inside the ‘ucf24’ folder created in step-1.
  6. Create a folder called ‘weights’ in the YOWO folder.
  7. Download pre-trained yolo weights, details given the GitHub page, and put it in the folder created in step-6.
  8. Download ‘resnext-101-kinetics.pth’ from the GitHub page and put in the ‘weights’ folder created in step -6.

For Validation

  1. Go to YOWO/evaluations/Object-Detection-Metrics
  2. Unzip the file ‘groundtruths_ucf.zip’
  3. Go back to the YOWO folder. Create a folder called ‘ucf_detections/detections_0’. Keep creating multiple subdirectories ‘detections_n’ for n = 0,1,2,3….. for further tasks.
  4. Run validation using the command

Run both training and validating model using the script: $ sh run_ucf101-24.sh

  1. Change the ‘detection_n’ portion in the script file for multiple validation tasks. python ./evaluation/Object-Detection-Metrics/pascalvoc.py --gtfolder groundtruths_ucf --detfolder ../../ucf_detections/detections_0

Thank you for your detailed guidance. I have successfully trained on ucf24 dataset. But trainning on ucf24 is more time consume than jhmdb21. What trouble me a lot is training on jhmdb21 dataset , since I find no bounding box ground truth file.

I have downloaded all jhmdb21 dataset, and didn't find training annotations just like ucf24's your provided. After searching on the internet, I didn't find exist bounding box annotation files.

To address this problem. I have generate bounding box ground truth files using jhmdb21 puppet mask files by myself. One .txt file for one per frame with contents : [label, xmin, ymin, xmax, ymax].

I have checked my groundtruth file with YOWO provided. In most cases they are similar to those provided，with a difference of one or two pixels. If anyone needed, you can download from jhmdb21_labels

Holmes-GU commented 4 years ago

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

yzfly commented 4 years ago

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

Holmes-GU commented 4 years ago

What about table two, recall and classification(2d + 3d + CFAM), can you reproduce them?

Zhihao Gu

gzhmsy@126.com | 签名由网易邮箱大师定制

On 07/9/2020 17:14，yzflynotifications@github.com wrote：

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Holmes-GU commented 4 years ago

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

When you valuate video_mAP for ucf-101, do you find 'testlist_video.txt'?

yzfly commented 4 years ago

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

When you valuate video_mAP for ucf-101, do you find 'testlist_video.txt'?

I generate it myself. In fact, I have given up my efforts on this project, too many questions and not a good baseline.

Holmes-GU commented 4 years ago

Okay, one more question, do you train the jhmdb-21 successfully? And do you have the gt for training?

Zhihao Gu

gzhmsy@126.com | 签名由网易邮箱大师定制

On 08/10/2020 17:39，yzflynotifications@github.com wrote：

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

When you valuate video_mAP for ucf-101, do you find 'testlist_video.txt'?

I generate it myself. In fact, I have given up my efforts on this project, too many questions and not a good baseline.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Riiick2011 commented 2 years ago

@yzfly Hi, have you reproduced the results in Table 2 of the paper?

My results are behind by 2-3 percentage points reported in the paper

Can you tell me what does your jhmdb experiment's cfg look like? My frame mAP is only about 59% for jhmdb. ----------jhmdb.yaml---------- TRAIN: DATASET: jhmdb21 # ava, ucf24 or jhmdb21 BATCH_SIZE: 24
TOTAL_BATCH_SIZE: 128 LEARNING_RATE: 1e-4 EVALUATE: False FINE_TUNE: False BEGIN_EPOCH: 1 END_EPOCH: 10 SOLVER: MOMENTUM: 0.9 WEIGHT_DECAY: 5e-4 STEPS: [3,4,5,6] LR_DECAY_RATE: 0.5 ANCHORS: [0.95878, 3.10197, 1.67204, 4.0040, 1.75482, 5.64937, 3.09299, 5.80857, 4.91803, 6.25225] NUM_ANCHORS: 5 OBJECT_SCALE: 5 NOOBJECT_SCALE: 1 CLASS_SCALE: 1 COORD_SCALE: 1 DATA: NUM_FRAMES: 16 SAMPLING_RATE: 1 TRAIN_JITTER_SCALES: [256, 320] TRAIN_CROP_SIZE: 224 TEST_CROP_SIZE: 224 MEAN: [0.4345, 0.4051, 0.3775] STD: [0.2768, 0.2713, 0.2737] MODEL: NUM_CLASSES: 21 BACKBONE_3D: resnext101 BACKBONE_2D: darknet WEIGHTS: BACKBONE_3D: "weights/resnext-101-kinetics-hmdb51_split1.pth" BACKBONE_2D: "weights/yolo.weights" FREEZE_BACKBONE_3D: True FREEZE_BACKBONE_2D: True LISTDATA: BASE_PTH: "/data1/su/datasets/JHMDB-YOWO" TRAIN_FILE: "/data1/su/datasets/JHMDB-YOWO/trainlist.txt" TEST_FILE: "/data1/su/datasets/JHMDB-YOWO/testlist.txt" TEST_VIDEO_FILE: "/data1/su/datasets/JHMDB-YOWO/testlist_video.txt" MAX_OBJS: 1 CLASS_NAMES: [ "brush_hair", "catch", "clap", "climb_stairs", "golf", "jump", "kick_ball", "pick", "pour", "pullup", "push", "run", "shoot_ball", "shoot_bow", "shoot_gun", "sit", "stand", "swing_baseball", "throw", "walk", "wave" ] BACKUP_DIR: "backup/jhmdb" RNG_SEED: 1
NUM_GPUS: 4 VISBLE_GPUS: '"0, 1, 2, 3"' GPUS_ID: [0, 1, 2, 3]