Detic / datasets /README.md
AK391
files
159f437
|
raw
history blame
6.51 kB
# Prepare datasets for Detic
The basic training of our model uses [LVIS](https://www.lvisdataset.org/) (which uses [COCO](https://cocodataset.org/) images) and [ImageNet-21K](https://www.image-net.org/download.php).
Some models are trained on [Conceptual Caption (CC3M)](https://ai.google.com/research/ConceptualCaptions/).
Optionally, we use [Objects365](https://www.objects365.org/) and [OpenImages (Challenge 2019 version)](https://storage.googleapis.com/openimages/web/challenge2019.html) for cross-dataset evaluation.
Before starting processing, please download the (selected) datasets from the official websites and place or sim-link them under `$Detic_ROOT/datasets/`.
```
$Detic_ROOT/datasets/
metadata/
lvis/
coco/
imagenet/
cc3m/
objects365/
oid/
```
`metadata/` is our preprocessed meta-data (included in the repo). See the below [section](#Metadata) for details.
Please follow the following instruction to pre-process individual datasets.
### COCO and LVIS
First, download COCO and LVIS data place them in the following way:
```
lvis/
lvis_v1_train.json
lvis_v1_val.json
coco/
train2017/
val2017/
annotations/
captions_train2017.json
instances_train2017.json
instances_val2017.json
```
Next, prepare the open-vocabulary LVIS training set using
```
python tools/remove_lvis_rare.py --ann datasets/lvis/lvis_v1_train.json
```
This will generate `datasets/lvis/lvis_v1_train_norare.json`.
### ImageNet-21K
The ImageNet-21K folder should look like:
```
imagenet/
ImageNet-21K/
n01593028.tar
n01593282.tar
...
```
We first unzip the overlapping classes of LVIS (we will directly work with the .tar file for the rest classes) and convert them into LVIS annotation format.
~~~
mkdir imagenet/annotations
python tools/unzip_imagenet_lvis.py --dst_path datasets/imagenet/ImageNet-LVIS
python tools/create_imagenetlvis_json.py --imagenet_path datasets/imagenet/ImageNet-LVIS --out_path datasets/imagenet/annotations/imagenet_lvis_image_info.json
~~~
This creates `datasets/imagenet/annotations/imagenet_lvis_image_info.json`.
[Optional] To train with all the 21K classes, run
~~~
python tools/get_imagenet_21k_full_tar_json.py
python tools/create_lvis_21k.py
~~~
This creates `datasets/imagenet/annotations/imagenet-21k_image_info_lvis-21k.json` and `datasets/lvis/lvis_v1_train_lvis-21k.json` (combined LVIS and ImageNet-21K classes in `categories`).
[Optional] To train on combined LVIS and COCO, run
~~~
python tools/merge_lvis_coco.py
~~~
This creates `datasets/lvis/lvis_v1_train+coco_mask.json`
### Conceptual Caption
Download the dataset from [this](https://ai.google.com/research/ConceptualCaptions/download) page and place them as:
```
cc3m/
GCC-training.tsv
```
Run the following command to download the images and convert the annotations to LVIS format (Note: download images takes long).
~~~
python tools/download_cc.py --ann datasets/cc3m/GCC-training.tsv --save_image_path datasets/cc3m/training/ --out_path datasets/cc3m/train_image_info.json
python tools/get_cc_tags.py
~~~
This creates `datasets/cc3m/train_image_info_tags.json`.
### Objects365
Download Objects365 (v2) from the website. We only need the validation set in this project:
```
objects365/
annotations/
zhiyuan_objv2_val.json
val/
images/
v1/
patch0/
...
patch15/
v2/
patch16/
...
patch49/
```
The original annotation has typos in the class names, we first fix them for our following use of language embeddings.
```
python tools/fix_o365_names.py --ann datasets/objects365/annotations/zhiyuan_objv2_val.json
```
This creates `datasets/objects365/zhiyuan_objv2_val_fixname.json`.
To train on Objects365, download the training images and use the command above. We note some images in the training annotation do not exist.
We use the following command to filter the missing images.
~~~
python tools/fix_0365_path.py
~~~
This creates `datasets/objects365/zhiyuan_objv2_train_fixname_fixmiss.json`.
### OpenImages
We followed the instructions in [UniDet](https://github.com/xingyizhou/UniDet/blob/master/projects/UniDet/unidet_docs/DATASETS.md#openimages) to convert the metadata for OpenImages.
The converted folder should look like
```
oid/
annotations/
oid_challenge_2019_train_bbox.json
oid_challenge_2019_val_expanded.json
images/
0/
1/
2/
...
```
### Open-vocabulary COCO
We first follow [OVR-CNN](https://github.com/alirezazareian/ovr-cnn/blob/master/ipynb/003.ipynb) to create the open-vocabulary COCO split. The converted files should be like
```
coco/
zero-shot/
instances_train2017_seen_2.json
instances_val2017_all_2.json
```
We further pre-process the annotation format for easier evaluation:
```
python tools/get_coco_zeroshot_oriorder.py --data_path datasets/coco/zero-shot/instances_train2017_seen_2.json
python tools/get_coco_zeroshot_oriorder.py --data_path datasets/coco/zero-shot/instances_val2017_all_2.json
```
Next, we preprocess the COCO caption data:
```
python tools/get_cc_tags.py --cc_ann datasets/coco/annotations/captions_train2017.json --out_path datasets/coco/captions_train2017_tags_allcaps.json --allcaps --convert_caption
```
This creates `datasets/coco/captions_train2017_tags_allcaps.json`.
### Metadata
```
metadata/
lvis_v1_train_cat_info.json
coco_clip_a+cname.npy
lvis_v1_clip_a+cname.npy
o365_clip_a+cnamefix.npy
oid_clip_a+cname.npy
imagenet_lvis_wnid.txt
Objects365_names_fix.csv
```
`lvis_v1_train_cat_info.json` is used by the Federated loss.
This is created by
~~~
python tools/get_lvis_cat_info.py --ann datasets/lvis/lvis_v1_train.json
~~~
`*_clip_a+cname.npy` is the pre-computed CLIP embeddings for each datasets.
They are created by (taking LVIS as an example)
~~~
python tools/dump_clip_features.py --ann datasets/lvis/lvis_v1_val.json --out_path metadata/lvis_v1_clip_a+cname.npy
~~~
Note we do not include the 21K class embeddings due to the large file size.
To create it, run
~~~
python tools/dump_clip_features.py --ann datasets/lvis/lvis_v1_val_lvis-21k.json --out_path datasets/metadata/lvis-21k_clip_a+cname.npy
~~~
`imagenet_lvis_wnid.txt` is the list of matched classes between ImageNet-21K and LVIS.
`Objects365_names_fix.csv` is our manual fix of the Objects365 names.