gkordo
commented
1 year ago @oliviernguyenquoc, many thanks for your question. This is a very inspiring work, and indeed the two methods are pretty close since both develop an SSL training process using a similar set of augmentations. Therefore, we have included the published version of this work, i.e. the VRL [31] method, in our comparison to the state-of-the-art in Table 1. Yet, there are some key differences between the two approaches:
- In VRL, the authors build based on SSL for representation learning to train the backbone network for feature extraction and an aggregation network that extracts clip-level representations. The similarity function is a variant of the handcrafted Chamfer Similarity tailored for hashed vectors, which is non-parametric. In contrast, in our work, we opt to use SSL for similarity learning, so we train a parametric similarity function without altering the representation space. To this end, our backbone network remains frozen during training, and we train a ViSiL-based network on top.
- Another thing is that, in VRL, they build a training dataset offline, which means that the video pairs used in training are static, so fixed variations are seen by the network. Instead, we dynamically generate the video pairs online during training; hence our model is exposed to a larger variety of combinations.
- Also, the research problem that VRL target to address is near-duplicate video retrieval. However, we tailor our method to tackle a wide range of similarity-based problems, including copy (near-duplicate), incident and event video retrieval, and their detection counterparts.
- Finally, as noted, we plan to release our implementation to the public, including the training processed and pretrained models. Additionally, we rely solely on publicly available datasets to build our models.
I am closing this issue since it is not directly related to this repo. Feel free to ask anything if there are further queries.
oliviernguyenquoc
https://www.semanticscholar.org/reader/93d3b14ac1c5b86dc7ea2eef54d4541e4e2df57f
This work seems to be very similar. Did you take inspiration of it ?
How do you think do you perform vs this work ?
If you release the code open source, it already make a huge difference.
Anyway, your paper is very insightful !