mmdet_sam

MMDetection-SAM

The current research direction of general-purpose object detection is moving toward large multimodality models. In addition to image inputs, most recent research outcomes incorporate text modality to improve the performance. Once the text modality is added, some very good properties of generic detection algorithms start to emerge, such as:

a large amount of easily accessible text data can be leveraged for joint training.
easy implementation of open-set object detection, leading to genuinely universal detection tasks.
can be used with the superb models in NLP to create some interesting and useful features.

Recently, Meta AI proposed Segment Anything which claims to be able to segment any object, and there are many great implementations based on it. MMDet integrates many high-performance and easy-to-use detection models, so why not we combine MMDet models and Segment Anything together to create something interesting?

From the current point of view, generic object detection can be divided into two main categories:

Closed-Set Object Detection, which can only detect a fixed number of classes of objects that appear in the training set
Open-Set object detection, which can also detect objects and categories that are not in the training set.

With the popularity of multimodal algorithms, open-set object detection has become a new research area, in which there are three popular directions:

Zero-Shot Object Detection, also known as zero-sample object detection, which emphasizes that the object categories of the testing set are not in the training set.
Open-Vocabulary Object Detection, which detects all objects with given category list that appears in the target images.
Grounding Object Detection, which predicts the location of the objects with given text descriptions that appears in the target images.

These three directions are not completely distinguishable in practice, but only different in general terms. Based on the above descriptions, we provide inference and evaluation scripts for multiple models to work with Segment Anything, which accomplished the following features:

supports classic Closed-Set object detection models in MMDet to work with SAM models for automatic detection and instance segmentation, such as Faster R-CNN and DINO.
supports Open-Vocabulary detection model like Detic to work with SAM models for automatic detection and instance segmentation.
supports Grounding Object Detection models to work with SAM models for automatic detection and instance segmentation, such as Grounding DINO and GLIP.
supports distributed detection and segmentation evaluation and automatic COCO JSON exportation across all models for user-friendly custom data assessment.

File Introduction

detector_sam_demo.py: for detection and instance segmentation on both single image and image folders.
coco_style_eval.py: for inference, evaluation and exportation on the given COCO JSON.
browse_coco_json.py: for visualizing exported COCO JSON.
images2coco.py: for customizd and unannotated COCO style JSON based on users' own image folder. This JSON can be used as the input to coco_style_eval.py.

This project referenced Grounded-Segment-Anything. Thanks!

Base Development Environment Setup

conda create -n mmdet-sam python=3.8 -y
conda activate mmdet-sam
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmengine

git clone https://github.com/open-mmlab/playground.git
cd playground

Feature Introduction

This project has included many outstanding open-sourced algorithms, in order to reduce the burden of the environment installation. If you prefer not to use a certain part of the features, you can skip the corresponding part. Our project can be divided into the following three sections.

1 Open-Vocabulary + SAM

Use Open-Vocabulary object detectors with SAM models. Currently we support Detic.

Dependencies Installation

pip install -U openmim
mim install "mmcv>=2.0.0"

# build from source
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection; pip install -e .; cd ..

pip install git+https://github.com/facebookresearch/segment-anything.git
pip install git+https://github.com/openai/CLIP.git

Demonstration

cd mmdet_sam

# download weights
mkdir ../models
wget -P ../models/ https://download.openmmlab.com/mmdetection/v3.0/detic/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth
wget -P ../models/ https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t cat \
    --sam-device cpu

The result will be generated at outputs/cat_remote.jpg like the following one:

We can also detect the remote by editing --test-prompt. Please be aware that you must use empty space and . to separate different categories.

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t "cat . remote" \
    --sam-device cpu

The generated outputs/cat_remote.jpg is now like this:

You can also run inferences on a folder by using this command:

# input a folder
python detector_sam_demo.py ../images \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t "cat . remote" \
    --sam-device cpu

The result images will be generated at the outputs folder.

If the graphics memory of your GPU can only support running one model, you can use --cpu-off-load to make sure every time there will be only one model running on the GPU:

# input a folder
python detector_sam_demo.py ../images \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t "cat . remote" \
    --cpu-off-load

We also support CPU inference by using --det-device cpu --sam-device cpu.

As Detic includes the mask results, we add a additional parameter --use-detic-mask. This allows us to run Detic only without running SAM models.

# input a folder
python detector_sam_demo.py ../images \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t "cat . remote" \
    --det-device cpu \
    --use-detic-mask

If you would like to visualize the results only, you can use set --only-det to run without SAM models.

# input a sinle image
python detector_sam_demo.py ../images/cat_remote.jpg \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t "cat" \
    --only-det

The outputs/cat_remote.jpg now is like this:

2 MMdet models + SAM

Use MMDet models with SAM models for instance segmentation tasks. Currently all MMDet models are supported.

Dependencies Installation

pip install -U openmim
mim install "mmcv>=2.0.0"

# build from source
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection; pip install -e .; cd ..

Demonstration

You can run all features like what we have covered in the above Detic section. The only difference is that you do not need to set --text-prompt. Here we demonstrate some classic usages.

Faster R-CNN models

cd mmdet_sam

mkdir ../models
wget -P ../models/ https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    ../mmdetection/configs/faster_rcnn/faster-rcnn_r50_fpn_2x_coco.py \
    ../models/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth \
    --sam-device cpu

DINO models

cd mmdet_sam

mkdir ../models
wget -P ../models/ https://download.openmmlab.com/mmdetection/v3.0/dino/dino-5scale_swin-l_8xb2-12e_coco/dino-5scale_swin-l_8xb2-12e_coco_20230228_072924-a654145f.pth

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    ../mmdetection/configs/dino/dino-5scale_swin-l_8xb2-12e_coco.py \
    dino-5scale_swin-l_8xb2-12e_coco_20230228_072924-a654145f.pth  \
    --sam-device cpu

3 Grounding models + SAM

Use Grounding object detectors with SAM models for instance segmentation tasks. Currently we support Grounding DINO and GLIP.

Dependencies Installation

Grounding DINO:

cd playground
pip install git+https://github.com/facebookresearch/segment-anything.git
pip install git+https://github.com/IDEA-Research/GroundingDINO.git # Please make sure your PyTorch, GCC and NVCC are all compatible to build CUDA ops successfully

GLIP:

cd playground

pip install git+https://github.com/facebookresearch/segment-anything.git
pip install einops shapely timm yacs tensorboardX ftfy prettytable pymongo transformers nltk inflect scipy pycocotools opencv-python matplotlib

git clone https://github.com/microsoft/GLIP.git
cd GLIP; python setup.py build develop --user # Please make sure your PyTorch, GCC and NVCC are all compatible to build CUDA ops successfully

Demonstration

Still, the usages are identical to the Detic part, and we only demonstrate part of the features here.

cd mmdet_sam

mkdir ../models
wget -P ../models/ https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    configs/GroundingDINO_SwinT_OGC.py \
    ../models/groundingdino_swint_ogc.pth \
    -t "cat . remote" \
    --sam-device cpu

The result will be generated at outputs/cat_remote.jpg like the following one:

cd mmdet_sam

mkdir ../models
wget -P ../models/ https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_a_tiny_o365.pth

# input a single image
python detector_sam_demo.py ../images/cat_remote.jpg \
    configs/glip_A_Swin_T_O365.yaml \
    ../models/glip_a_tiny_o365.pth \
    -t "cat . remote" \
    --sam-device cpu

4 COCO JSON evaluation

We support running coco_style_eval.py in both a distributed and a non-distributed manner. By default, this script runs on COCO Val2017 dataset organized in the format shown as follows:

├── ${COCO_DATA_ROOT}
│   ├── annotations
│      ├──── instances_val2017.json
│   ├── val2017

We use Detic to demonstrate how it works here, other algorithms are same to this.

cd mmdet_sam

# non-distributed
python coco_style_eval.py ${COCO_DATA_ROOT} \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t coco_cls_name.txt

# distributed on eight cards on one machine
bash ./dist_coco_style_eval.sh 8 ${COCO_DATA_ROOT} \
    configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py \
    ../models/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth \
    -t coco_cls_name.txt

5 COCO evaluation results

Method	bbox thresh	Test set	Box AP	Seg AP
Detic	0.2	COCO2017 Val	0.465	0.388
Detic	0.001	COCO2017 Val	0.481	0.403
GroundingDino	0.3	COCO2017 Val	0.419
GroundingDino	0.0	COCO2017 Val	0.469
GroundingDino*	0.3	COCO2017 Val	0.404
GLIP	0.0	COCO2017 Val	0.429

Note: *means use original GroundingDINO approach to test

6 Custom dataset

We use a example to illustrate how a custom dataset can be used to obtain an annotation file using GroundingDINO and SAM.

Data prepare

Use the following command to download the cat dataset:

cd playground

wget https://download.openmmlab.com/mmyolo/data/cat_dataset.zip
mkdir data
unzip cat_dataset.zip -d data/cat
rm cat_dataset.zip

Note:，Need to replace 1 cat with cat in cat/class_with_id.txt

Use the images2coco.py script to generate the unlabeled json file:

cd mmdet_sam
python images2coco.py ../data/cat/images ../data/cat/class_with_id.txt cat_coco.json

Inference

Here we use GroundingDINO and SAM model as an example for inference, and get the prediction json file:

python coco_style_eval.py ../data/cat/ \
      configs/GroundingDINO_SwinT_OGC.py \
      ../models/groundingdino_swint_ogc.pth \
      -t ../data/cat/class_with_id.txt \
      --data-prefix images \
      --ann-file annotations/cat_coco.json \
      --out-dir ../cat_pred \
      --sam-device cpu

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

mmdet_sam

mmdet_sam

README.md

MMDetection-SAM

File Introduction

Base Development Environment Setup

Feature Introduction

1 Open-Vocabulary + SAM

Dependencies Installation

Demonstration

2 MMdet models + SAM

Dependencies Installation

Demonstration

3 Grounding models + SAM

Dependencies Installation

Demonstration

4 COCO JSON evaluation

5 COCO evaluation results

6 Custom dataset

Data prepare

Inference

Name		Name	Last commit message	Last commit date
parent directory ..
configs		configs
README.md		README.md
README_zh-CN.md		README_zh-CN.md
browse_coco_json.py		browse_coco_json.py
coco_cls_name.txt		coco_cls_name.txt
coco_style_eval.py		coco_style_eval.py
detector_sam_demo.py		detector_sam_demo.py
dist_coco_style_eval.sh		dist_coco_style_eval.sh
images2coco.py		images2coco.py
predictor_glip.py		predictor_glip.py
utils.py		utils.py

Files

mmdet_sam

Directory actions

More options

Directory actions

More options

Latest commit

History

mmdet_sam

Folders and files

parent directory

README.md

MMDetection-SAM

File Introduction

Base Development Environment Setup

Feature Introduction

1 Open-Vocabulary + SAM

Dependencies Installation

Demonstration

2 MMdet models + SAM

Dependencies Installation

Demonstration

3 Grounding models + SAM

Dependencies Installation

Demonstration

4 COCO JSON evaluation

5 COCO evaluation results

6 Custom dataset

Data prepare

Inference