Skip to content

Latest commit

 

History

History
140 lines (123 loc) · 5.44 KB

README.md

File metadata and controls

140 lines (123 loc) · 5.44 KB
Raw

Creating ActivityNet SRL (ASRL) from ActivityNet Captions (AC) and ActivityNet Entities (AE)

The code is for generating the dataset from the parent datasets. If you just want to use use ASRL as a training bed, you can skip this. See data

Quick summary

Very briefly, the process is as follows:

  1. Add semantic roles to captions in AC.
  2. Prepocess AE. In particular, resize all the proposals, ground-truth bounding boxes (this is required for SPAT/TEMP).
  3. Preprocess the features and choose only 5 groundtruths for GT5 setting.
  4. Obtain the bounding boxes and category names from AE for the relevant phrases.
  5. Filter out some verbs like "is", "are", "complete", "begin"
  6. Filter some SRL Arguments based on Frequency.
  7. Get Training/Validation/Test videos.
  8. Do Contrastive Sampling and store the dictionary files for easier sampling during training.

Preprocessing Steps

  1. First download relevant files. Optional: specify the data folder where it would be downloaded.

    bash download_asrl_parent_ann.sh [save_point]

    The folder should look like:

    anet_cap_ent_files
|-- anet_captions_all_splits.json (AC captions)
|-- anet_entities_test_1.json
|-- anet_entities_test_2.json
|-- anet_entities_val_1.json
|-- anet_entities_val_2.json
|-- cap_anet_trainval.json        (AE Train annotations)
|-- dic_anet.json                 (Train/Valid/Test video splits for AE)

  2. Use SRL Labeling system from AllenAI (Should take ~15 mins) to add the semantic roles to the captions from AC.

    cd $ROOT
python dcode/sem_role_labeller.py

    This will create $ROOT/cache_dir and store the output SRL files which should look like:

    cache_dir/
|-- SRL_Anet
|-- SRL_Anet_bert_cap_annots.csv  # AC annotations in csv format to input into BERT
|-- srl_bert_preds.pkl            # BERT outputs

  3. Resize the boxes in AE.

    cd $ROOT
python dcode/preproc_anet_files.py --task='resize_boxes_ae'

    This takes the file cap_anet_trainval.json as input (this is the main AE annotation file) and outputs anet_ent_cls_bbox_trainval.json. The latter file contains resized ground-truth boxes. It also resizes the proposal boxes, taking in anet_detection_vg_fc6_feat_100rois.h5 as input and produces anet_detection_vg_fc6_feat_100rois_resized.h5 as output. The latter contains resized proposals.

  4. GT5 setting

    cd $ROOT
python dcode/preproc_anet_files.py --task='choose_gt_5'

    Intially, there are 100 proposals per frame. For faster iteration, we only choose the 5 proposals from each frame. If there is a ground-truth box, we take include that box, and the remaining are included in order of their proposal score (not a fair way, but the best that could be done). If there are no ground-turth box, we choose the top5 scoring proposals.

    To compute the recall scores (for sanity check):

    python dcode/preproc_anet_files.py --task='compute_recall'

    By default, it computes recall scores for GT5, you can change the proposal file, for other settings.

  5. Aligning SRL outputs and NounPhrases from AE to create ASRL and adding the bounding boxes to the ASRL files (<1min)

    cd $ROOT
python dcode/asrl_creator.py

    Now $ROOT/data/anet_srl_files/ should look like:

    anet_srl_files/
|-- verb_ent_file.csv        # main file with SRLs, BBoxes
|-- verb_lemma_dict.json     # dictionary of verbs corresponding to their lemma

  6. Use the Train/Val videos from AE to create Train/Val/Test videos for ASRL (~5-7 mins). Additionally, create the vocab file for the SRL arguments

    cd $ROOT
python dcode/prepoc_ds_files.py

    This will create anet_cap_ent_files/csv_dir. It should look like:

    csv_dir
|-- train.csv
|-- train_postproc.csv
|-- val.csv
|-- val_postproc.csv

    Further, now $ROOT/data/anet_srl_files/ should look like:

    anet_srl_files/
|-- trn_verb_ent_file.csv       # train file
|-- val_verb_ent_file.csv       # val & test file
|-- verb_ent_file.csv
|-- verb_lemma_dict.json

  7. Do Constrastive sampling for train and validation set (~30mins)

    cd $ROOT
python code/contrastive_sampling.py

    Now your anet_srl_files directory should look like:

    anet_srl_files/
|-- trn_asrl_annots.csv			# used for training
|-- trn_srl_obj_to_index_dict.json		# used for CS
|-- trn_verb_ent_file.csv			# not used anymore
|-- val_asrl_annots.csv			# used for val/test
|-- val_srl_obj_to_index_dict.json		# used for CS
|-- val_verb_ent_file.csv			# not used anymore
|-- verb_ent_file.csv			# not used anymore
|-- verb_lemma_dict.json			# not used anymore

  8. I have provided drive links to the processed files (generated after completing all the previous steps):

    1. anet_cap_ent_files and anet_srl_files: https://drive.google.com/open?id=1mH8TyVPU4w7864Hxiukzg8dnqPIyBuE3
    2. SRL_Anet: https://drive.google.com/open?id=1vGgqc8_-ZBk3ExNroRP-On7ArWN-d8du
    3. resized proposal h5 files: https://drive.google.com/open?id=1a6UOK90Epz7n-dncKAeFDQP4TBgqdTS9
    4. fc6_feats_5rois: https://drive.google.com/open?id=13tvBIEAgv4VS5dqkZBK1gvTI_Z22gRLM

    Alternatively, you can download these files from download_asrl_parent_ann.sh by passing asrl_proc_files:

    bash download_asrl_parent_ann.sh asrl_proc_files