search

Na-Z / sess

[CVPR2020 Oral] SESS: Self-Ensembling Semi-Supervised 3D Object Detection
MIT License
138 stars 29 forks source link

would it boost performence using non-labeled val data? #8

Closed turboxin closed 4 years ago

turboxin commented 4 years ago

Hi @Na-Z，thanks for sharing this great work.

According to tabel2 of the paper, SESS outperforms Votenet with a big margin with 100% training label, I'm wondering did SESS use non-labeled val data for semi-surpervised training? If not, would it boost performence using non-labeled val data?

Looking forward for your reply

Na-Z commented 4 years ago

Hi, the result of SESS in Table 2 did not use any data from val set during training.

Based on the observations in Table 1, the performance of SESS would be further improved with additional unlabeled data (e.g. those unlabeled data from val data).

Hope this answers your question. :)

turboxin commented 4 years ago

Thank you for the quick reply~