search

WongKinYiu / yolor

implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks (https://arxiv.org/abs/2105.04206)
GNU General Public License v3.0
1.99k stars 517 forks source link

what should be the ideal mAP0.95 should be? #276

Open saumya221 opened 2 years ago

saumya221 commented 2 years ago

Hi, I am training yolor_csp with a single class of ~1500 images, where 400 images are negative images. train test split is of the ratio 80:20. While training with fine-tune hyper parameters for 1280 size for 5000 epochs, best_overall is being generated around epoch #900. after than mAP values are going down. The best mAP0.5:0.95 is coming around 0.85. My query is , is that value too high ? or is it expected? Am I overfitting the data ? Shall I continue training beyond 5000 epochs? Kindly give your valuable feedback.

SanKumSan commented 2 years ago

What do you mean by Negative Samples ?

saumya221 commented 2 years ago

Negative sample means images not having the object of interest.