Figure 2.2: Appearance of beluga whales in different backgrounds.

Figure 2.2 (a)

Figure 2.2 (b)

Figure 2.2 (c)

Figure 2.2 (d)

Figure 2.3: Appearance of Narwhal whales in different backgrounds.

Figure 2.3 (a)

Figure 2.3 (b)

Figure 2.7: Patches from size 256 to 4096. Orientation of the waves is critical to discern whales from waves.

Figure 2.14 : Prediction of FasterRCNN with ConvNext-T backbone on various patch sizes. Increasing the patch size helps the model eliminate false positives due to waves.

Figure 2.14 (a) - Predictions of FasterRCNN-ConvNext-256

Figure 2.14 (b) - Predictions of FasterRCNN-ConvNext-512

Figure 2.14 (c) - Predictions of FasterRCNN-ConvNext-1024

Figure 2.14 (d) - Predictions of FasterRCNN-ConvNext-2048

Figure 2.14 (e) - Predictions of FasterRCNN-ConvNext-4096

Figure 2.15 : Prediction of FasterRCNN with ConvNext-T backbone on various backgrounds.

Figure 2.15 (a)

Figure 2.15 (b)

Figure 2.15 (c)

Figure 2.15 (d)

Figure 2.16 : Predictions on images overlooked by DFO but found by our best model.

Figure 2.16 (a) - contains around 2000 whales

Figure 2.16 (b) - contains around 865 whale

Figure 3.3 : The small ice shards in 2023 surveys closely resemble the beluga whales of the 2014 survey.

Figure 3.3 (a) - Model prediction on 2023 survey

Figure 3.3 (b) - Beluga whales in 2014 dataset

Figure 3.4: Bowhead whale which is not present in the earlier surveys missed by our model

Figure 3.5: Beluga whales detected with low confidence along with some clear whales not predicted.

Figure 3.12: Images with fewer than five objects acquired using our acquisition function.

Figure 3.12 (a) shows the model’s predictions before fine-tuning.

Figure 3.12 (b) shows the predictions after finetuning.

Figure 3.13: Image with more than 20 objects acquired using our acquisition function.

Figure 3.13 (a) shows the model’s predictions before fine-tuning.

Figure 3.13 (b) shows the predictions after finetuning.