This repo contains my trained models/logs for Kaggle Competition - TF Barrier Reef Challenge

Used Ultralytics/Yolov5 weights to re-train on Challenge Dataset.

NOTE 1: all weights of training are on my Kaggle account private dataset.

NOTE 2: weights for "s" model are added, as they are under 100mb

NOTE 3: all training is done on Kaggle Free GPU quota of 36 hours/week

NOTE 4: only the best submission data is logged here

Final Results

• Winner Rank : 1 out of 2026, Score : 0.760
• My Final Rank : 1307 out of 2026, Score : 0.572

Leaderboard Scores

• [13 Feb 2022] 1st Competitor : 0.806 || My Score : 0.520
• [23 Jan 2022] 1st Competitor : 0.779 || My Score : 0.520
• [17 Jan 2022] 1st Competitor : 0.720 || My Score : 0.520
• [15 Jan 2022] 1st Competitor : 0.692 || My Score : 0.520
• [02 Jan 2022] 1st Competitor : 0.673 || My Score : 0.513
• [02 Jan 2022] 1st Competitor : 0.672 || My Score : 0.513
• [30 Dec 2021] 1st Competitor : 0.658 || My Score : 0.512
• [26 Dec 2021] 1st Competitor : 0.658 || My Score : 0.502
• [20 Dec 2021] 1st Competitor : 0.619 || My Score : 0.488
• [19 Dec 2021] 1st Competitor : 0.619 || My Score : 0.473
• [16 Dec 2021] 1st Competitor : 0.619 || My Score : 0.325

Challenge Submission Score Log

• First submission

  • Date : 14/Dec/2021
  • Model : Yolov5l
  • Dataset : default, original with annotation
  • My score : 0.325
  • Device : GPU
  • Leaderboard Rank #1 score : 0.619

• Second submission

  • Summary : Model Ensembling, with 1st model running on original images and 2nd running on enhanced images with different Conf + IOU thresholds for each model. Results of both models are fused using Weighted-Boxes-Fusion
  • Date : 19/Dec/2021
  • Model : Yolov5l (previous) + Yolov5l (retrained on enchanced images at 1280)
  • Dataset : Original + Enhanced Images
  • My score : 0.473
  • Device : GPU
  • Leaderboard Rank #1 score : 0.619

• Third submission

  • Summary : Model Ensembling with "Test Time Augment" set to false and threshold tuned, 3 models are used. Results of both models are fused using Weighted-Boxes-Fusion
  • Date : 20/Dec/2021
  • Model : Yolov5l (previous) + Yolov5l (retrained on enchanced images at 1280) + Yolov5l (retrained on enchanced images at 1280)
  • Dataset : Original + Enhanced Images
  • My score : 0.488
  • Device : GPU
  • Leaderboard Rank #1 score : 0.619

• Fourth submission

  • Summary : Model ensembling + new image pre-processing
  • Date : 26/Dec/2021
  • Model : Yolov5l (previous) + Yolov5l6 (retrained on new CLAHE processed images at 1280)x2
  • Dataset : New Enhanced Images
  • My score : 0.502
  • Device : GPU
  • Leaderboard Rank #1 score : 0.658

• Fifth submission

  • Same as previous with params modifications.
  • Date : 30/Dec/2021
  • My score : 0.512

• Sixth submission

  • Param modifications
  • Date : 26/Jan/2022
  • My score : 0.520

Model Train Log

• "yolov5l train 1"

  • Model trained : Yolov5l
  • Trained on Image Size : 1024
  • Epochs : 15
  • Dataset details : Trained on original data. Only images with annotations were used. No image preprocessing is used.

• "yolov5x6 train 1" (This is a saved checkpoint)

  • Model trained : Yolov5x6
  • Trained on Image Size : 1024
  • Epochs : 7
  • Dataset details : Trained on enhanced images. Images with/without annotations were used.

• "yolov5l train 2"

  • Model trained : Yolov5l (train1 weights)
  • Trained on Image Size : 1280
  • Epochs : 12
  • Dataset details : Trained on ehanced+augmented data.

• "yolov5l6 train"

  • Model trained : Yolov5l6
  • Trained on Image Size : 1280
  • Epochs : -
  • Dataset details : Trained on new ehanced+augmented data.

• "yolov5s6 train"

  • Model trained : Yolov5s6
  • Trained on Image Size : 1280
  • Epochs : 40
  • Dataset details : Trained on new ehanced+augmented data.