DBAI DL Challenge

Welcome to DBAI challenge

The goal of the challenge is to train model that would work well on Indian roads.

Content :

Repo consist of mAP script for evaluation of the model with training data and test data in data directory

<Tree structure>

dl_challenge
├── Readme.md
├── compute_AP.py                 |  # mAP calculation script
└── data                          |
    ├── test                      |
    │   └── dbai_test_data.zip    |  # test data images
    └── train                     |
        ├── dbai_train_data.zip   |  # coco style training data
        └── gt.pickle             |  # ground truth pickle file

• compute_AP.py :Script to compute AP result for evaluation of model

# usage 

GT_PICKLE = "path/to/_GT_/pickle_file"
PRD_PICKLE = ""path/to/_PRD_/pickle_file""

compute_mAP(GT_PICKLE,PRD_PICKLE)

# Output mAP values are saved as csv file at the <prd>.pickle file path 

Pickle file format:

Bellow is the format for predicted.pickle file which will then be used for calculating mAP score using compute_mAp script.

[
    [image_name,
            [ 
                [class_id , class_name , confidence , [x1 , y1 , x2 , y2]
                [class_id , class_name , confidence , [x1 , y1 , x2 , y2]
                ....
            ]
     ]
    ...
]

Deliverable

Send us the following details at hr@drivebuddyai.co

• Predicted pickle file for given test dataset

• Inference in 10 random samples from training data

• Inference in 10 random samples from test data

• Code file or project (training and inference code)

• Training log

• Document containing details of

  • Model used
  • Hyper parameter and setting used for training
  • Setting used for testing / inference
  • Training log
  • Approach
  • Your observations and conclusion

Evaluation

Candidate will be evaluated by

• mAP score per class
• Approach and model selection