search

VisualComputingInstitute / triplet-reid

Code for reproducing the results of our "In Defense of the Triplet Loss for Person Re-Identification" paper.
https://arxiv.org/abs/1703.07737
MIT License
764 stars 216 forks source link

Performance on different dataset not good #59

Closed SmitSheth closed 6 years ago

SmitSheth commented 6 years ago

Hello, I used the pre-trained model given in this repository on my own image dataset to calculate embeddings and then calculated eucledian distance between the embeddings. Testing on same person with different view, postures, gives bad result(eucledian distance is high for some images and low for some). Link to sample images of a person i tested on https://drive.google.com/open?id=15QKkosOP6sRFB8xzJz8O36OXEvEfegeI.

NOTE: I have used the pre-trained model given.

Pandoro commented 6 years ago

Sadly the model is not super robust to such variants because they did not appear in the training data. While the framework itself is by no means the issue here, you will need to gather additional training data, or create a mix of multiple datasets in order to train a more general model that might be robust to such inputs.

We will likely not be doing such a thing in in the near future. So I'll close this issue. If you train such a model and have additional results or questions feel free to re-open it.