BDD Project

Link to Analysis: Analysis

Prerequisites

Ensure you have Docker installed on your system. Pull the repo and run

docker build -t bdd-project-env

I tried finetuning RTDETR for 6 epochs: Checkpoint Please put it under train/checkpoints/

Workflow Steps

1. Create DataFrames

Run the script to create dataframes for analysis:

docker run --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to local repo folder>:/bdd_files" \
    bdd-project-env \
    python data/create_dataframes.py \
    --base-path /data \
    --output-path /bdd_files

2. Run Streamlit for Analysis

Start the Streamlit app to perform analysis:
PLEASE SCROLL HORIZONTALLY AND COPY THE ENTIRE COMMAND, STREAMLIT ACTS WEIRD WITH MULTILINE ARGPARSE HENCE I HAD TO PUT IT ALL IN ONE LINE

docker run -p 8501:8501 --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    streamlit run data/dashboard_with_val.py -- --train-df-path /bdd_files/extracted_data.pq --val-df-path /bdd_files/extracted_data_val.pq --base-data-path /data

3. Create COCO Files for Training

Generate the necessary COCO files:

docker run --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    python data/json_to_coco.py \
    --bdd-json-path /data/bdd100k_labels_release/bdd100k/labels/bdd100k_labels_images_train.json \
    --output-json-path /data/bdd100k_images_100k/bdd100k/images/100k/train/_annotations.coco.json \
    --split-name train

docker run --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    python data/json_to_coco.py \
    --bdd-json-path /data/bdd100k_labels_release/bdd100k/labels/bdd100k_labels_images_val.json \
    --output-json-path /data/bdd100k_images_100k/bdd100k/images/100k/val/_annotations.coco.json \
    --split-name val

4. Train the Model

Train the RT-DETR model:

docker run --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    python train/train_rtdetr.py

5. Evaluate the Model

Evaluate the trained model and generate metrics:

docker run --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    python train/eval_model.py --output-path /bdd_files

6. Visualize Results

Run the Streamlit app to visualize results and perform qualitative analysis:
PLEASE SCROLL HORIZONTALLY AND COPY THE ENTIRE COMMAND, STREAMLIT ACTS WEIRD WITH MULTILINE ARGPARSE HENCE I HAD TO PUT IT ALL IN ONE LINE

docker run -p 8501:8501 --rm \
  -v "/path/to/assignment_data_bdd/:/data" \
  -v "<path to where you pulled the repo>:/bdd_files" \
    bdd-project-env \
    streamlit run eval/eval_dashboard.py -- --matched-predictions-path /bdd_files/results_iou.json --image-dir /data/bdd100k_images_100k/bdd100k/images/100k/val --processed-data-path /bdd_files/extracted_data_val.pq --coco-annotations-path /data/bdd100k_images_100k/bdd100k/images/100k/val/_annotations.coco.json

Notes

• Ensure all paths are correctly set according to your project structure.
• Adjust the Docker volumes as necessary for your environment.