# Does redudancy in dataset effect the model's generalization

My current dataset consists of 8000 images along with their corresponding labels which denote the object's coordinates.

I am trying to implement a model that predicts the object's bounding box and class.However,I noticed that several of the images in the dataset tend to have the same bounding box coordinates. For example for a particular class say X there are 300 images,and out of those 300 images,50% of the images,tend to have the same bounding box coordinate.

As such when I trained my model on this dataset,the model was unable to predict the bounding box correctly ,whereas when I trained the same model on standard dataset like COCO,the bounding box predictions were accurate.

As such ,I would like to know ,is the redundancy in the dataset be one factor that is affecting the model performance.

Sample examples from the dataset are shown below

Image 1 =>label values [0.23,0.45,0.33,0.6]
Image 2=>label values [0.23,0.45,0.33,0.6]
Image 3=>label values [0.23,0.45,0.33,0.6]
.
.
.
.
.
Image 20=>label values [0.23,0.45,0.33,0.6]

All images are same except the object in each image has been translated


Any insight or suggestions on whether my reasoning is correct ,will be helpful.