Object Detection and Object captioning - examples created using published model

This fork adds a Google Colab link to try model
Evalution notes of overall scene summary generation using GRiT + Detic with ChatGPT
Original repo - below

Dense captioning of images - use case 1

Object Detection - use case 2

Object Detection - behaviour of incorrectly mapping objects to other known classes - this is due to the absence of the classes in the training set - not an inherent deficiency of the model

Notes on using GRiT+Detic with ChatGPT

Captioning for detected objects with GRiT captures some aspect of the object features beyond just the object class name. This is a consequence of the captioning component getting object features in the form of image patches. For instance in some pictures, the captioning describes the "sky" (detected object) as clear vs dark etc.
This particular aspect is perhaps a distinguising factor of GRiT's approach compared to using a model like Detic which only outputs object classes. Detic's object detection capability may appear to be better than GRiT in images tested only because of the difference in training set. So if we combine these two model outputs and use them as input to ChatGPT we get a rich overall scene summary. Images were selected from a royalty free site Pexels.

Related Links

Twitter thread showing the performance of GRiT+Detic with ChatGPT. ChatGPT output quality improves for Winograd schemas when passing it bounding box information as well as dense captioning output by GRiT

Twitter thread comparing this model to Detic with ChatGPT. GRiT's dense captioning output improves ChatGPT's narrative compared to Detic which only detects classes with bounding boxes

A Hugging Face app 🤗 using to Detic to output object classes with coordinates and dimensions which can be used to cut-and-paste to ChatGPT playground for dense captioning.

A Generative Region-to-text Transformer for Object Understanding

GRiT is a general and open-set object understanding framework that localizes objects and describes them with any style of free-form texts it was trained with, e.g., class names, descriptive sentences (including object attributes, actions, counts and many more).

GRiT: A Generative Region-to-text Transformer for Object Understanding
Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, Lijuan Wang
¹State University of New York at Buffalo, ²Microsoft
arXiv technical report (PDF)

Installation

Please follow Installation instructions.

Object Understanding Demo - One Model Two tasks

Download the GRiT model or use the following commend to download:

mkdir models && cd models
wget https://datarelease.blob.core.windows.net/grit/models/grit_b_densecap_objectdet.pth && cd ..

The downloaded GRiT model was jointly trained on dense captioning task and object detection task. With the same trained model, it can output both rich descriptive sentences and short class names by varying the flag --test-task. Play it as follows! 🤩🤩🤩

Output for Dense Captioning (rich descriptive sentences)

python demo.py --test-task DenseCap --config-file configs/GRiT_B_DenseCap_ObjectDet.yaml  --input demo_images --output visualization --opts MODEL.WEIGHTS models/grit_b_densecap_objectdet.pth

Output for Object Detection (short class names)

python demo.py --test-task ObjectDet --config-file configs/GRiT_B_DenseCap_ObjectDet.yaml  --input demo_images --output visualization --opts MODEL.WEIGHTS models/grit_b_densecap_objectdet.pth

Output images will be saved under the visualization folder, which looks like:

Benchmark Inference and Evaluation

Please follow dataset preparation instructions to download datasets.

Download our trained models and put them to models/ for evaluation.

Object Detection on COCO 2017 Dataset

Model	val AP	test-dev AP	Download
GRiT (ViT-B)	53.7	53.8	model
GRiT (ViT-L)	56.4	56.6	model
GRiT (ViT-H)	60.4	60.4	model

To evaluate the trained GRiT on coco 2017 val, run:

# GRiT (ViT-B)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet --eval-only MODEL.WEIGHTS models/grit_b_objectdet.pth
# GRiT (ViT-L)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_L_ObjectDet.yaml --output-dir-name ./output/grit_l_objectdet --eval-only MODEL.WEIGHTS models/grit_l_objectdet.pth
# GRiT (ViT-H)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_H_ObjectDet.yaml --output-dir-name ./output/grit_h_objectdet --eval-only MODEL.WEIGHTS models/grit_h_objectdet.pth

Dense Captioning on VG Dataset

Model	mAP	Download
GRiT (ViT-B)	15.5	model

To test on VG test set, run:

python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_DenseCap.yaml --output-dir-name ./output/grit_b_densecap --eval-only MODEL.WEIGHTS models/grit_b_densecap.pth

It will save the inference results to output/grit_b_densecap/vg_instances_results.json. We use the VG dense captioning official evaluation codebase to report the results. We didn't integrate the evaluation code into our project as it was written in Lua. To evaluate on VG, please follow the original codebase's instructions and test based upon it. We're happy to discuss in our issue section about the issues you may encounter when using their code.

Training

To save training memory, we use DeepSpeed for training which can work well for activation checkpointing in distributed training.

To train on single machine node, run:

python train_deepspeed.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet

To train on multiple machine nodes, run:

python train_deepspeed.py --num-machines 4 --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet

Acknowledgement

Our code is in part based on Detic, CenterNet2, detectron2, GIT, and transformers. We thank the authors and appreciate their great works!

Citation

If you find our work interesting and would like to cite it, please use the following BibTeX entry.

@article{wu2022grit,
  title={GRiT: A Generative Region-to-text Transformer for Object Understanding},
  author={Wu, Jialian and Wang, Jianfeng and Yang, Zhengyuan and Gan, Zhe and Liu, Zicheng and Yuan, Junsong and Wang, Lijuan},
  journal={arXiv preprint arXiv:2212.00280},
  year={2022}
}

taskswithcode / grit Goto Github PK

grit's Introduction

Object Detection and Object captioning - examples created using published model

Table of contents