Revision history [back]

yea, that's expected. cascades are for rigid things with reliable texture and pose.

that's not, what you have.

you probably tried to train it on "aquarium fish", but that's a far too broad category, and any variation in pose or texture will degrade the detection

I have 183 positive, 270 negative images.

even if it were the appropriate method, you would need like 50x the data

to train something successful,

• you need a lot more data
• a different mechanism (e.g. retrain a detection cnn, like yolo or ssd)
• more refined categories / object classes