Episode 2 for my VIPrior challenge journey link!

Issue 1

Using mmclassification / Creating Object Detection Backbone

Goal: Improving Object Detection Performance

1. Generate datasets by cropping images using bounding boxes (using roboflow)
2. Train an image classification pretrained model (image encoder) using mmclassification
3. Load the well-trained image encoder and use it for training our detection model

Tool: mmclassification

To ensure compatibility with our current setting (mmdetection), planning to use mmclassification for training.

• mmclassification
• mmcls docx

---

Issue 2

Using Roboflow

Data Versioning

• Goal: Pretrain the image encoder to improve performance.
• Approach: Crop images within the bounding boxes and train for classification.
• Method: Go to "Versions" in roboflow -> Select the dataset -> Choose "isolate objects" in preprocessing -> then Dataset is generated!
• Result: A new version of the dataset has been created.

Old Dataset

New Dataset