yd-yin / fishermatch Goto Github PK

Can you tell me the details about visualizing the Fisher distribution on SO(3)? As it is in Figure 1.

Thanks for amazing work! But I didn't find the public weights file, like: .ckpt .tar. Do we have the pretrained models for testing or evaluation? Thanks for download linking~

Hi, thanks for your work. It's very interesting. I realized that during supervision part, the loss function is different between the code and paper, right? In the paper, the loss function for supervision is simple (negative log likelihood of the groundtruth rotation in the predicted distributions). Why the loss function in the code is different? Does this loss function (in the code) have been mentioned in the paper?

Thanks!

Hi, nice to see your interesting work. But I am curious about the entropy-based filtering part.
There seems to be a strong correlation between the Fisher entropy [equation 9] $H(A)$ and the negative log-likelihood of the mode w.r.t. its parameter: $-\log (p(R^{\text{mode}}|A))$, where $R^{\text{mode}}=\text{procrustes}(A)$. The negative log-likelihood can also show the prediction confidence. So why not perform filtering by the likelihood?

Dear author, thank you for your excellent work! I have a question about the code in train.py. Why we use the best_median_error for filtering and choosing the best model? What if we use the best_mean_error as an indicator.

Hi! Thank you for your inspiring work!

I noticed that the entropy threshold $\tau$ is set to -5.3. Why is it negative? Shouldn't entropy always positive?

@yd-yin
Hi, Yingda:

Could you mind sharing the way to get averaged results of 6 categories in Pascal3D+ dataset? I have trained each category and computed average result following the pascal.yml setting (ss_ratio: 20). However, I cannot get the same results with your paper's results.
My reproduced results:

Your results in Table2: