This is not the official repository for CDMnet (pdf) or DMnet (pdf)

• There are two challenges in detecting objects from aerial images:
  • Most objects appear small
  • Uneven distribution of object positions reduces detection efficiency
• CDMnet and DMnet is designed to detect small objects more efficiently in urban aerial scenes, and we need to detect floating objects on the sea, both of which have the same characteristics of being small in comparison to the scene

An Overview of our framework

There are already state of art algorithms that can achieve satisfying results on density map generation. In our project, the density map generation modular uses MCNN to achieve the task.

We introduce code from Ma to train MCNN.

The pretrain weight can be accessed here.

There are a lot of state-of-art real-time object detector that can achieve fast and high accuracy object detection nowadays. In our project, we prefer Yolov7-tiny as the detector.

We introduce code from WongKinYiu to train Yolo

Here provided the finetuned models on Marine Debris that available for direct application in tiny marine debris detection.

• Due to license problem with our dataset, it is not available for download.
• Prepare your dataset with point labels (for training MCNN) and annotation labels (for training yolo)

• Run the following command to install required library

pip install -r requirements.txt

• We introduce code from Ma to train MCNN.
• We introduce code from WongKinYiu to train Yolo
• Put your MCNN param file under MCNN_weights/ and Yolov7-tiny weights file under Yolov7_weights/

• Prepare your images and run the command using pretrained MCNN and Yolov7-tiny for jpg image

python detect.py --img_dir samples/Trash.jpg --mcnn_param MCNN_weights/mcnn_marine_debris.param --yolo_weights Yolov7_weights/best.pt --half

• The ideas of our implementation is come from CDMnet and DMnet
• Density map generation code is borrowed from the MCNN-pytorch
• Detector code is borrowed from the Official YOLOv7
• Thanks for their nice contribution.