Official repository of the paper "The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding".
The benchmarks described in the paper can be found in the benchmarks
folder of this repository. Refer to Creates dataset to find the steps for reproducing the benchmark generation process.
Additionally, curated training and validation sets, created following the benchmark generation procedure, are available in the training_sets
and validation_sets
folders.
- 🔥 09/2024: The extension of this work, "Is CLIP the main roadblock for fine-grained open-world perception?", won the Best Paper Award at CBMI 2024!
- 🔥 04/2024: "The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding" has been selected as a highlight at CVPR2024, representing just 11.9% of accepted papers.
- 🔥 04/2024: An extension of this work, "Is CLIP the main roadblock for fine-grained open-world perception?", is now available in pre-print (arXiv) (Code).
- 🔥 03/2024: The first FG-OVD trainining sets are available in this repository at
training_sets
. - 🔥 02/2024: "The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding" has been accepted to CVPR2024!
To perform the dataset collection it will be necessary to create a Docker container using the following commands:
git clone https://github.com/lorebianchi98/FG-OVD.git
cd FG-OVD/captions_generation/docker
docker image build -t IMAGE_NAME - < Dockerfile
To use OpenAssistant LLaMa-based (the model used in our experiment in the official experiments is OpenAssistant LLaMa 30B SFT 6) it is necessary to have access to LLaMa 30B by Meta AI and to obtain OpenAssistant weight following the guidelines provided here. The model should be placed in captions_generations/models.
Retrieve the PACO dataset and place the desired JSON file for processing into the captions_generations/datasets directory. In our case, we utilized paco_lvis_v1_test.json.
To create the whole dataset by manipulating PACO json and interacting with OpenAssistant.
python main.py --gpu DEVICE_ID --dataset paco_lvis_v1_test --batch_size BATCH_SIZE
We used a batch size of 4. This command will create a benchmark with 10 temporary hardnegatives of Hard type. To create the hardnegatives of Hard, Medium, Easy, Trivial, Color, Material and Transparency type, with 10 hardnegatives, it is necessary to run the following commands:
cd negative_generations
./creates_datasets.sh ../jsons ../OUT_DIR 10
The dataset follows the standard LVIS format:
data["images"]: # a list of dictionaries, each dictionary corresponds to one image
{
'id': int,
'file_name': str,
'width': int,
'height': int,
'license': int,
'coco_url': str,
'flickr_url': str
}
data['annotations']: # a list of dictionaries, each dictionary correspond to one annotation
{
'id': int,
'bbox': [x,y,width,height],
'area': float,
'category_id': int,
'image_id': int,
'segmentation': RLE,
'neg_category_ids': int, # not on LVIS
}
data["categories"]: # a list of dictionaries, each dictionary corresponds to one object category
{
'id': int,
'name': str,
'def': str, # always ''
'image_count': int,
'instance_count': int,
'synset': str, # always ''
'synonyms': List(str), # always []
'frequency': char, # always 'f'
}
The mAP in the paper for Detic with the number of negatives equal to 2 (N=2) depicted in Table 3 for the Transparency benchmark is 24.6 instead of 28. Thanks to iMayuqi for reporting this issue.
If you found this code useful, please cite the following paper:
@inproceedings{bianchi2024devil,
title={The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding},
author={Bianchi, Lorenzo and Carrara, Fabio and Messina, Nicola and Gennaro, Claudio and Falchi, Fabrizio},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={22520--22529},
year={2024}
}