add backgroud class in box_nms

[arcadiaphy]

Description

This PR is mentioned in #14057

What I have done in box_nms operator:

1. add background_id argument
2. filter out background boxes before sorting operation

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Feature1, tests, (and when applicable, API doc)
• Feature2, tests, (and when applicable, API doc)

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

[vandanavk]

@mxnet-label-bot add [pr-awaiting-review, Operator]

[ChaiBapchya]

Retrigger the CI as the CI issue of test_stn has been fixed 2 days back by #14063

git commit --allow-empty -m "Trigger notification"

[ChaiBapchya]

curious to know why this change?

[arcadiaphy]

I think the valid id_index should be >= 0 instead of just > 0. Surely, this line is not relevant to this PR.

[ChaiBapchya]

@mxnet-label-bot update [pr-awaiting-response, Operator]

[ankkhedia]

@apeforest @samskalicky Coul d you please help review this PR?

[anirudhacharya]

@arcadiaphy please add tests for these changes.

[wkcn]

Excellent. Thanks for your contribution!
LGTM: )

[arcadiaphy]

@wkcn CI seems OK now, ready for merge.

[zhreshold]

would marking scores of all boxes with background_id to 0 or -1 simpler and don't require any operator change?

[arcadiaphy]

@zhreshold Currently there are no background_id logic in box_nms operator on prefiltering or nms step, so a lot of time is wasted in processing background boxes.

To keep operator unchanged, background boxes need to be removed in NDArray input.

[zhreshold]

@arcadiaphy Yes, a pre-filtering can be much easier, but there's no guarantee that box_nms will always have clean input. So I am actually convinced this is good to go. BTW, do you have the new validation time with this new PR?

[arcadiaphy]

@zhreshold Sometimes the validation time of SSD_512 on voc dataset is more than half an hour before, much longer than training time.

I have written a simple script to show benchmark:

import mxnet as mx
import numpy as np
import time

np.random.seed(0)
ctx = mx.gpu(0)

def test_nms(background_ratio):
    batch_size = 32
    prior_number = 100000  # SSD_512 has more than 100k prior boxes
    background_number = int(prior_number * background_ratio)
    data = np.zeros((batch_size, prior_number, 6))
    data[:, :, 0] = 0
    data[:, :background_number, 0] = -1
    data[:, :, 1] = np.random.random((batch_size, prior_number))

    xmin = np.random.random((batch_size, prior_number))
    ymin = np.random.random((batch_size, prior_number))
    width = np.random.random((batch_size, prior_number))
    height = np.random.random((batch_size, prior_number))
    data[:, :, 2] = xmin
    data[:, :, 3] = ymin
    data[:, :, 4] = xmin + width
    data[:, :, 5] = ymin + height

    mx_data = mx.nd.array(data, ctx=ctx)
    start = time.time()
    rv = mx.nd.contrib.box_nms(mx_data, overlap_thresh=0.5, valid_thresh=0, topk=400, id_index=0, score_index=1, coord_start=2, background_id=-1)
    mx.nd.waitall()
    print 'background ratio {} process time: {} s'.format(background_ratio, time.time() - start)


test_nms(0)
test_nms(0.5)
test_nms(0.9)
test_nms(0.99)

The output on TITAN X (Pascal):

background ratio 0 process time: 22.8679060936 s
background ratio 0.5 process time: 10.5633630753 s
background ratio 0.9 process time: 1.97785997391 s
background ratio 0.99 process time: 0.20063495636 s

[zhreshold]

@arcadiaphy Please also take a look at two improvements contributed by @ptrendx
#14352 #14290

[zhreshold]

waiting for justification of the added API

[arcadiaphy]

@zhreshold My PR is complementary to the great improvements of @ptrendx on cuda kernels. After combining all three PRs, box_nms is close to the original MultiBoxDetection operator.

[zhreshold]

@arcadiaphy Please resolve the conflict and it's good to go. 😺

[arcadiaphy]

@zhreshold OK to go now.

[zhreshold]

thanks @arcadiaphy , this is merged