Coordinate-Aware Mask R-CNN with Group Normalization: Towards Improved Underwater Instance Segmentation

We propose a novel hybrid Instance Segmentation (IS) method, called Coordinate-Aware Mask R-CNN (CAM-RCNN), which is motivated by the Mask R-CNN (MRCNN) and Segmenting Objects by LOcation v2 (SOLOv2) algorithms. CAM-RCNN’s structure is mainly constructed on MRCNN with modifications replicated from SOLOv2’s architecture. To this end, we extend MRCNN by introducing a CoordConv layer, multiple Group Normalization (GN) data normalisation layers, and the unique compound DiceBCE Loss (DBL) function to its mask prediction branch. We also present a box-based version of SOLOv2’s mask-based Matrix Non-Maximum Suppression (MNMS), termed Matrix Bounding Box Non-Maximum Suppression (MBBNMS), which replaces the original Batched NMS (BNMS) of MRCNN applied to its predicted region proposals. For training and evaluation, we adopt several real-world unique aquatic datasets provided by the University of Aberdeen. Along with the aforementioned models, we consider two other State-Of-The-Art (SOTA) techniques, namely CenterMask and Conditional Convolutions for Instance Segmentation (CondInst), to better validate our findings. Our results indicate that CAM-RCNN achieves an outstanding test set generalisation performance compared to the other SOTA architectures on almost all evaluation metrics considered. Particularly, the finalised CAM-RCNN model demonstrates over 2.5× (177.3%) and 31.9% increase in instance mask Average Precision (AP) regarding the worst and best performing baseline models, respectively. Our proposed CAM-RCNN is promising to be a generic approach for IS-related applications.

Model	AP	AP$\mathbf{_{50}}$	AP$\mathbf{_{75}}$	AP$\mathbf{^{bb}}$	AP$\mathbf{_{50}^{bb}}$	AP$\mathbf{_{75}^{bb}}$
CenterMask	14.9	19.8	18.2	15.9	19.6	18.8
CondInst	8.8	11.2	10.9	8.9	11.1	10.8
MRCNN	18.5	24.9	22.6	18.8	24.4	22.8
SOLOv2	14.5	19.8	16.7	-	-	-
CAM-RCNN	24.4	31.5	30.2	24.2	31.3	27.5

Table 1. Quantitative results of instance mask and Bounding Box (BB) of different methods in the target domain. $\mathbf{bb}$ denotes BBOX.

Installation

1. Install the latest conda (Anaconda).

2. Download the source code and extract to the desired location.

3. Enter an Anaconda Prompt terminal window in administrator mode.

4. Navigate to the source code directory.

5. Create the aquatic (default) conda environment by running:

   conda env create -f environment.yml
   

• Alternatively, to change the default conda environment name use:

  conda env create -n <env_name> -f environment.yml
  

  where <env_name> should be replaced with the new custom name of the environment.

6. Activate the project conda environment via:

   conda activate <env_name>
   

   where <env_name> is as above.

7. Run a script file (dataset and COCO pretrained model weights are needed).

• demo.py - Script for creating demo inference videos
• inspect_results.py - Script used for inspecting loss results along with validation and test set output metrics of a model
• plot_dist.py - Script plotting the data distribution across the training, validation and test sets
• run_amsrcr.py - Script that runs the AMSRCR image enhancement technique on a given image dataset
• train_eval_model.py - Script used for training an instance segmentation model
• visualize_json_results.py - Script used for visualising model JSON IS results (i.e. coco_instances_results.json)

Manuals

• Maintenance manual
• User manual