Have you fine-tune the Reid model on MOT dataset?
Opened this issue · 27 comments
Hi @longcw , I've replaced the Reid part of extracting features by my own trained model, the evaluation result gets worse. Have your fine-tune the Reid model on the MOT dataset?
No, the reid model was trained on a combination of market1501 and duke reid dataset.
I implemented the part-aligned reid (https://arxiv.org/abs/1707.07256) and got about 79.7 rank-1 on market 1501 dataset. Then the trained model was used for tracking without finetuning.
I found that the order of magnitudes of our Reid feature are different, maybe I need to normalize the Reid feature
Hi @longcw , after I normalize my Reid feature, I found the distribution of dists changes from approximately [0.1, 1.5] to [0.003, 0.08] (my feature size is 1x2048, and the original is 1x512). I think I only need to modify the min_ap_dist value and I actually tried some(0.05,0.06, etc.), but the result's always worse than the original ones. It's strange because I didn't change any other part and my Reid model is tested to be stronger on Market1501(90.0% rank-1), do you have any reminder on this point?
I am wondering why the distance changed to [0.003, 0.08] after you normalized the feature. L2 distance on two normalized features is same to the cosine distance (only different from some constants). So it should be in [0, 1] or something similar.
Second, I also tested resnet50 model with about 90% rank-1 on market1501 while the tracking performance is similar or even worse. I tried to train the model on market1501 and test it on duke and only got about 20%-30% rank-1. Reid models are very easy to overfit on the training set. So I train the model on three datasets to achieve better generation ability. Worse reid performance on the training set doesn't imply a worse tracking result. I think this is one reason you got a worse result.
To improve the model for tracking, the first thing is to improve the generation ability rather than achieve a higher score on the training set, unless you train the model on the tracking dataset directly.
Thank you, you've got a great point. I think the distance range still in [0,1], but the numbers are really small. I gonna train the Reid model on more dataset and let you know the results~
market1501, duke, and cuhk03.
But I am still wondering why you got the range of [0, 0.2]. It's an unnormal small range for cosine distance or L2 distance on normalized features.
For example, the following image is the distance distribution on MOT16-02.
I didn't normalize the feature in training, that's the only reason I can think about...
I change the value of the min_ap_dist and visualize the result on MOT17, it seems not bad...
Hi, I tried with a reid model using normalized feature in training, the distance value is still small.
While, as I add more data into training set, the distance value gets larger, e.g. [0,0.1] for model on Market1501, [0, 0.2] for model on Market1501+Duke, [0, 0.3] for model on Market1501+Duke+Msmt17.
Maybe it's because of the network structure? For my model, I use resnet50 as backbone and get the conv5_x feature for further classification, while your model split the feature and align the feature with body part, so your feature map differs more in each part?
Maybe you can try to plot the distribution of positive and negative distances as I showed before, to find the best threshold and check the range.
So I have a question，how to train the reid model for MOT dataset or others? fine-tune the trained Reid model(we have trained on Market1501,Duke etc)on the MOT dataset or train on MOT dataset from other imgnet pretrain model? @longcw
@longcw I plot the cosine distance of MOT-02 with my Reid model:
why do you have three different histograms in your image?
@ I'm trying to reproduce your reid model and I notice you define 'SpatialCrossMapLRN' in googleNet, what is this used for?
@Kewenjing1020 @ygren The reid model was trained by my pytorch re-implementation of https://github.com/zlmzju/part_reid.
Maybe I can share the training code later if you are interested in it.
@longcw Your work is very interesting！I'm very interested in your training process.Looking forward to your sharing so that I can reproduce your results.Thx！
Hi @longcw, @Kewenjing1020 can you tell me how to re-train Patchclassifier from scratch?
@shibu38 Hi, I didn't try to reproduce patch classifier yet.
@longcw I tried to retrained reid by adding a classifier after your network and fine-tune from your model, but the code always corrupts. Then I removed the LRN layer, then it runs well but with a really bad result...
@Kewenjing1020 Hi, do you have any idea how to do that?? or Do you have any plan in future to that??
@longcw
Hi, my features distance finally become normal. It's because the image preprocess part doesn't consist with that in my model training code.
Shame of me taking so long to find the bug. TT
Is googleNet for Reid? And squeezeNet for what?
Are they both from https://github.com/zlmzju/part_reid?
I generally know that you get the re-id model from https://github.com/zlmzju/part_reid which uses part-aligned reid and GoogleNet. But I still do not understand why you need two models and what are their functions?
@longcw Can you solve my question? Thanks!
I generally know that you get the re-id model from https://github.com/zlmzju/part_reid which uses part-aligned reid and GoogleNet. But I still do not understand why you need two models and what are their functions?
Section 3.2, 3.4 here
@Crazy-stones: As the author mentioned, reading the paper might help.