@fengyuentau, when it's expected to be merged? I believe, the patch is very useful.

@fengyuentau, when it's expected to be merged? I believe, the patch is very useful.

We can, but since block quantization in opencv does not have any acceleration in terms of inference speed, merging this patch only increases the size of zoo and we are not gaining any benifit from it for now. Merging is postponed until block quantization works practically better in opencv.

I don't quite get it. We don't see any performancedegradation with block-wise quantized modelsand the models get smaller, not bigger. That is, we get smaller size + same speed. But if you are still in doubt, let's discuss it next week at our meeting, pls, include it into the agenda

I don't quite get it. We don't see any performancedegradation with block-wise quantized modelsand the models get smaller, not bigger. That is, we get smaller size + same speed. But if you are still in doubt, let's discuss it next week at our meeting, pls, include it into the agenda

Okay, lets do it in the next meeting.

Just to add my two cents: despite not achieving an inference speed improvement, this PR significantly reduces the network size while retaining the original accuracy.

Further optimization could be done in the future, also including adapting block-wise quantization for the new inference engine.

@DaniAffCH Thank you for all the effort on this pull request. We decided to merge this pull request with the following changes:

1. Could you report the accuracy of text detection and recoginition as well? Even though they may not have a good number, we still want to know about this since it is here in the pull request.
2. Change model naming suffix from_blocked to_int8bq. Add a note in eachREADME explaining models with suffix_int8bq is block-quantized in int8 precision.

Also need to provide each command that you use to generate the block-quantized models. Just want to ensure reproducibility.

@DaniAffCH Do you plan to push commits to finalize this PR? If no, I will merge this one first then do it in the subsequent PRs.

Yes I'll definitely address your comments. I've been busy with other projects, but now I can finalize this PR.

I've just updated the file suffixes and the related README files.

Could you report the accuracy of text detection and recoginition as well? Even though they may not have a good number, we still want to know about this since it is here in the pull request.

Regarding text detection, I couldn't find a suitable evaluation script ineval, so I don't know how to test it.

Regarding text recognition, I decided not to include the English version because of a severe drop in accuracy. Such a drop in accuracy doesn't occur in the Chinese version, so I decided to include onlytext_recognition_CRNN_CHtext_recognition_CRNN_CN.
However, the evaluation script only supports IIIT5K and ICDAR2003 datasets containing English words, so it's not possible to test the Chinese models on them.

All the models have been block-quantized usingthe block quantization script with the following command:

python block_quantize.py --input_model INPUT_PATH --block_size 64

Done!