Not working when testing on our own image dataset! [without training]

[ashishmokalkar]

Hello, I used the pre-trained model given in this repository on my own image dataset to calculate embeddings and then calculated co-sine similarity between the embeddings. Testing on same person with different view, postures, I am getting very less co-sine similarity. But, testing on person with different view of CUHK dataset, the co-sine similarity is very high.

NOTE:

• I have used the pre-trained model given.
• I have not performed any training on my dataset.

[Pandoro]

Hello @ashishmokalkar ,

we have never really tested the generalization, but I know it can generalize to other data pretty well as long as the data has similar "features". Is the camera quality similar, are the crops the same, do the poses look similar as those in Market-1501, etc.

However, even ignoring the dataset bias, computing the cosine similarity using embeddings from our model doesn't make any sense. It was trained using an euclidean distance and while that is equivalent to the cosine similarity when the embeddings are normalized, this is not the case for our embeddings.

So don't use the cosine similarity, use a euclidean distance instead and even if that doesn't work, there is no guarantee of generalization. CNNs are very good at overfitting certain details in data which might not be present in other datasets. Without seeing your data I don't have any feeling for how good or how bad this is supposed to work.