The BDD100k dataset is a large dataset of object detection annotations. This repo includes three different notebooks. A notebook for preprocessing the labels, a notebook for exploratory analysis and a notebook for modeling.
The BDD100k inclues a single large json file for train and a single json file for val labels. In order to not have to load the large json files into memory, we preprocess the labels into a multiple json files.
Exploring the dataset includes different histograms of the categories of objects, the number of objects per image, and the number of objects per category.
These notebooks include finetuning on a Faster R-CNN model. The most import part of the modeling is the customer dataset class which is used by pytorch dataloader. The target object is a dict with a few different keys: boxes, labels, and image_id are a few examples. The boxes key is a list of bounding boxes and their coordinates (x1, y1, x2, y2). The labels key is a list of objects detected per image.
Model evaluation is done by using the official pytorch evaluation script (engine.py). The best mAP@0.5:0.95 is 24.1%.