It is always better for the model to see the object in as many different contexts as possible, this helps it learn robustly. So the images should ideally contain the objects taken from different angles, under varying illuminations and at different distances from the camera
2. Medium Resolution
High-resolution images take much longer to process and may capture unnecessary details while a low resolution will not contain any features for the model to learn. It is best to find a balance between the two depending on the size of your object. Best to resize your images so that the object is bigger than 32 x 32 pixels. For the free tier, we have a maximum size of ~1000 x 1000 for your images.
3. Large dataset
More the number of images with the objects that the model sees, the better it learns