1. The final output of the problem is a binary classification problem.

2. Initially I found out the cosine similarity between consecutive shots.

3. As the four features place,act,cast,audio are distinct so I wanted to fit a nonline boundary. I chose KNN and RandomForest classifier.

4. Since there is impbalance in datasets I went ahead with sampling techniques and as expected Oversampling performed well compared to undersampling.

5. Check install.md for package installations

6. Check Ml_imbalance.ipynb for results and analysis

7. Download the given dataset and store it in data folder

8. Run final.py --datain "./data" --dataout "./out/" for the best found model which fits and evaluates the model. Pass data input folder and output folder location.

9. Run evaluate_sceneseg.py --datadir "./out/" for getting the evaluation results

    "AP": 0.4418872028438688,
    "mAP": 0.4564401595678161,
    "Miou": 0.45414800530021754,
    "Precision": 0.2761656092479825,
    "Recall": 0.7473442326299846,
    "F1": 0.39309552999275693
}

Train Test split 0.75/0.25 Test accuracy is : 0.9235327420267975

Scores: {
    "AP": 0.30079042370763087,
    "mAP": 0.3106786272526226,
    "Miou": 0.3426166826535341,
    "Precision": 0.5342522938717706,
    "Recall": 0.12604315828627943,
    "F1": 0.19920247085457896
}

Test accuracy is : 0.7355727495753915

Scores: {
    "AP": 0.5812168468330086,
    "mAP": 0.5735931851621817,
    "Miou": 0.4793941305957929,
    "Precision": 0.276096372587527,
    "Recall": 0.8623173522748953,
    "F1": 0.4087102162264981
}

Test accuracy is : 0.6485752028684657

Scores: {
    "AP": 0.17974113103317563,
    "mAP": 0.19112205712931624,
    "Miou": 0.37837250328419675,
    "Precision": 0.16354380841309596,
    "Recall": 0.7612370742132829,
    "F1": 0.26256056850200127
}

Test accuracy is : 0.6843177958105303

Scores: {
    "AP": 0.2634912915127264,
    "mAP": 0.27305655507348414,
    "Miou": 0.43162796376145063,
    "Precision": 0.21485679610619718,
    "Recall": 0.8768767300525099,
    "F1": 0.33727663186040563
}

The Best Performing model is KNN with oversampling with features corresponding to cosine similarity between adjancent features of the shots.

This Baseline 2 model is better than the given coarse prediction results

Test accuracy is : 0.9522537909168971

Scores: {
    "AP": 0.8231448030014414,
    "mAP": 0.8183934846996804,
    "Miou": 0.7100984502390212,
    "Precision": 0.6279891510330122,
    "Recall": 0.8546849823773823,
    "F1": 0.7161831235608485
}

This Performs better than KNN

Test accuract is : 0.7470843555387809

Scores: {
    "AP": 0.7264627480785202,
    "mAP": 0.7146891065555759,
    "Miou": 0.6663586410503086,
    "Precision": 0.5494326852458175,
    "Recall": 0.8637291228154672,
    "F1": 0.663004900009928
}

Default parameters for Random forest produce good results

Random Forest Classifier model could fit the data well and produce non-linear decision boundaries for this scene classification task