issues about visualization

[captain0305]

Hi!

I have some problem with visualization using model model_final_resnet_bdd.

using code

Results are showing the det boxes on the img of all id classes. But there's no ood boxes showing on the img. How could I tell apart the id and old object using this repo?

Looking forward to your reply!

[captain0305]

I have solve this issue with this code

but I still have some question with this issue.

inter_feat = results.inter_feat[0:max_boxes]
    print(inter_feat)
    print(torch.logsumexp(inter_feat[:, :-1], dim=1).cpu().data.numpy())
    print((np.argwhere(
            torch.logsumexp(inter_feat[:, :-1], dim=1).cpu().data.numpy() < energy_threshold)).reshape(-1))
    if energy_threshold:
        labels[(np.argwhere(
            torch.logsumexp(inter_feat[:, :-1], dim=1).cpu().data.numpy() < energy_threshold)).reshape(-1)] = 8

What is the inter_feat? It has 11 dimension. And after calculating , it is being used to tell apart the ood and id obj. Is this just match the step in the paper below?

And I wonder why inter_feat has 11 dimension and the meaning of each dimension. And also why the pred_cls_probs has 11 dimension? I just using the model model_final_resnet_bdd, and it's config just have 8 categories. I just quite confused about that.

Looking forward to your reply

[d12306]

hi, @captain0305 , sorry for the late reply! i am really busy recently with experiments.

1) yes, the inter_feat here is the logits.

2) actually bdd100k has 10 classes plus a background class. here we use the 11-bit logit for training but essentially we only train with 8 classes. this is the legacy code used in https://github.com/xinw1012/cycle-confusion. you can definitely use 8-bit. i suppose the result to be similar.

[d12306]

@captain0305 , feel free to post more questions here if needed! i will check more frequently from now on

[captain0305]

Thank u for your reply! I have trained my data using stud, and have some benchmarks.

[BigBuffa1o]

@captain0305 Hello,by applying your code on a single test img,should I modify the original yaml file to get a successful output like you?