This is a Tensorflow 2.3 implementation of the paper YOLACT: Real-time Instance Segmentation and YOLACT++: Better Real-time Instance Segmentation. The paper presents a fully-convolutional model for real- time instance segmentation that achieves 29.8 mAP on MS COCO at 33.5 fps evaluated on a single Titan Xp, which is significantly faster than any previous competitive approach. This Repo implements "ResNet50-FPN". Unlike original implemetation of YOLACT/YOLACT++ in which image is resized to 550x550, this repo can handle image of size MxN.

• MaskIOU loss added.
• DCNv2 added.

• Protobuf 3.0.0
• Tensorflow (>=2.3.0), Tensorflow Addons
• cocoapi
• OpenCV

For detailed steps to install Tensorflow, follow the Tensorflow installation instructions. A typical user can install Tensorflow using one of the following commands:

pip install tensorflow==2.3

pip install tensorflow-gpu==2.3

The remaining libraries can be installed on Ubuntu 16.04 using via apt-get:

sudo apt-get install protobuf-compiler

Download the cocoapi. The default metrics are based on those used in Pascal VOC evaluation.

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
make
make install

Protobufs is used to configure model and training parameters. Before the framework can be used, the Protobuf libraries must be compiled. This should be done by running the following command from the yolact/ directory:

# From yolact/
protoc protos/*.proto --python_out=.

Note: If you're getting errors while compiling, you might be using an incompatible protobuf compiler. If that's the case, use the following manual installation

If you are on linux:

Download and install the 3.0 release of protoc, then unzip the file.

# From yolact/
wget -O protobuf.zip https://github.com/google/protobuf/releases/download/v3.0.0/protoc-3.0.0-linux-x86_64.zip
unzip protobuf.zip

Run the compilation process again, but use the downloaded version of protoc

# From yolact/
./bin/protoc protos/*.proto --python_out=.

If you are on MacOS:

If you have homebrew, download and install the protobuf with brew install protobuf

Alternately, run:

PROTOC_ZIP=protoc-3.3.0-osx-x86_64.zip
curl -OL https://github.com/google/protobuf/releases/download/v3.3.0/$PROTOC_ZIP
sudo unzip -o $PROTOC_ZIP -d /usr/local bin/protoc
rm -f $PROTOC_ZIP

Run the compilation process again:

# From yolact/
protoc protos/*.proto --python_out=.

1. Git clone https://github.com/tensorflow/addons.
2. Apply the patch named deformable_conv2d.patch.
3. Compile tensorflow addon. For example for cuda 10.1

# Only CUDA 10.1 Update 1 
cd addons
export TF_NEED_CUDA="1"

# Set these if the below defaults are different on your system
export TF_CUDA_VERSION="10.1"
export TF_CUDNN_VERSION="7"
export CUDA_TOOLKIT_PATH="/usr/local/cuda"
export CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu"

# This script links project with TensorFlow dependency
python3 ./configure.py

bazel build build_pip_pkg --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
bazel-bin/build_pip_pkg artifacts

4. Pass -use_dcn True and -use_mask_iou True to train.py

I've compiled tensorflow_addons-0.11.2(cuda 10.1, cudnn 7.6, tf 2.3.0), tensorflow_addons-0.13.0(cuda 11.2, cudnn 8.2, tf 2.5.0) and tensorflow_addons-0.17.0(cuda 11.6, cudnn 8.3, tf 2.6.2) for python3.6.

Note: While compiling tensorflow addon for tf 2.5+, change the header in addons/tensorflow_addons/custom_ops/layers/cc/kernels/deformable_conv2d_op.h from

#include "tensorflow/core/kernels/batch_matmul_op_impl.h"

to this:

#include "tensorflow/core/kernels/matmul_op_impl.h"

Also, with cuda 11.4+ where following error is coming:

status: Internal: too many resources requested for launch

need to reduce threads per block in deformable_conv2d.patch. For this change every occurrence of config.thread_per_block with config.thread_per_block/2.

Refer to the tensorflow object detection api for tfrecord creation. (link)

Each dataset is required to have a label map associated with it. This label map defines a mapping from string class names to integer class Ids. The label map should be a StringIntLabelMap text protobuf. Sample label maps can be found in object_detection/data. Label maps should always start from id 1. For an example:

item {
  id: 1
  name: 'Cat'
}


item {
  id: 2
  name: 'Dog'
}

For every example in your dataset, you should have the following information:

1. An RGB image for the dataset encoded as jpeg or png.
2. A list of bounding boxes for the image. Each bounding box should contain:
   1. A bounding box coordinates (with origin in top left corner) defined by 4 floating point numbers [ymin, xmin, ymax, xmax]. Note that we store the normalized coordinates (x / width, y / height) in the TFRecord dataset.
   2. The class id of the object in the bounding box.
3. PNG encoded mask for every groundtruth bounding box. Each mask has only a single channel, and the pixel values are either 0 (background) or 1 (object mask).

Training procedure can be conducted directly by following command:

python train.py -tfrecord_train_dir 'path of TFRecord training files'
                -tfrecord_val_dir 'path of TFRecord validation files'
                -pretrained_checkpoints 'path to pre-trained checkpoints (if any)'
                -label_map 'path label_map.pbtxt'
                -train_iter 'number of  iteration for training'
                -img_h 'image height'
                -img_w 'image width'
                -num_class 'No of classes excluding background'
                -aspect_ratio 'aspect ratio for anchors'
                -scale 'scales in pixels for anchors '
                -batch_size 'batch_size'
                -lr 'learning rate'
                -momentum 'momentum for SGD'
                -weight_decay 'weight_decay rate for SGD'
                -print_interval 'interval for printing training result'
                -save_interval 'interval for conducting validation'
                -valid_iter 'number of iteration for validation'

Inside saved_models there will be saved graphs according to the score of their validation. For an example saved_model_0.19968511 is saved_model when the validation loss was 0.19968511. To run inference on using this saved_model see infer.py.

• https://github.com/feiyuhuahuo/Yolact_minimal
• https://github.com/dbolya/yolact
• https://github.com/leohsuofnthu/Tensorflow-YOLACT/