Detection and Tracking Methods:

• Detection: I use a Faster R-CNN model for fish detection. This model is chosen for its accuracy and robustness in detecting objects in images.

• Tracking: I use the Deep SORT algorithm to track detected fish across video frames. Deep SORT is effective in maintaining object identities even after occlusions.

• ======================================================================

• Layer (type:depth-idx) Output Shape Param #

• ============================================================

• FasterRCNN [0, 4] --

• ├─FasterRCNN: 1-1 [0, 4] --

• │ └─GeneralizedRCNNTransform: 2-1 [2, 3, 320, 576] --

• │ └─BackboneWithFPN: 2-2 [2, 256, 5, 9] --

• │ │ └─IntermediateLayerGetter: 3-1 [2, 960, 10, 18] 2,947,552

• │ │ └─FeaturePyramidNetwork: 3-2 [2, 256, 5, 9] 1,467,392

• │ └─RegionProposalNetwork: 2-3 [0, 4] --

• │ │ └─RPNHead: 3-3 [2, 15, 10, 18] 609,355

• │ │ └─AnchorGenerator: 3-4 [6075, 4] --

• │ └─RoIHeads: 2-4 [0, 4] --

• │ │ └─MultiScaleRoIAlign: 3-5 [0, 256, 7, 7] --

• │ │ └─TwoMLPHead: 3-6 [0, 1024] 13,895,680

• │ │ └─FastRCNNPredictor: 3-7 [0, 2] 10,250

• ==========================================================

• Total params: 18,930,229

• Trainable params: 18,871,333

• Non-trainable params: 58,896

• Total mult-adds (G): 2.59

• =========================================================

• Input size (MB): 49.77

• Forward/backward pass size (MB): 263.78

• Params size (MB): 75.72

• Estimated Total Size (MB): 389.26

• ======================================================

Optimizations:

• Detection: The Faster R-CNN model is fine-tuned on a relevant dataset to improve detection accuracy.
• Tracking: Deep SORT parameters are adjusted to handle the specific characteristics of fish movement.
• Image Processing: CLAHE is used to enhance image contrast

training setups:

• epoches=10
• learning rate=1e-4
• batch size=1