search

tyomj / product_detection

Working with scale: 2nd place solution to Product Detection in Densely Packed Scenes
Apache License 2.0
48 stars 9 forks source link

readme

Working with scale: 2nd place solution to Product Detection in Densely Packed Scenes

Introduction

This repository contains code for the 2nd place solution of the detection challenge which is held within CVPR 2020 Retail-Vision workshop. For more information see my report. For all the experiments MMDetection v1 was used.

Dataset

The dataset has been originally announced by Eran Goldman et. al. In order to obtain the dataset for research purpose, please concat the authors.

Getting started

For evaluation purpose please clone pycocotools, change the parameter maxDets to 300 here and then install locally.

1. Convert SKU110k csv format to COCO-like json

python sku110k_scripts/sku110k_to_coco.py --args

2. Convert a full frame COCO-like dataset to a tiled one

python sku110k_scripts/split_on_tiles.py --args

3. Training with mmdet

./tools/dist_train configs/sku110k/sku110k_faster_rcnn_r50_fpn_anchor_1x_4tiles_test_half_res.py 2

4. Testing with mmdet

./tools/dist_test configs/sku110k/sku110k_faster_rcnn_r50_fpn_anchor_1x_4tiles_test_half_res.py workdir/faster_rcnn_r50_fpn_anchor_1x_4tiles/latest.pth 2 --eval bbox

5. Create a dummy json file for the leaderboard-test

python sku110k_scripts/lb_test_to_coco.py --args

6. Inferencing with mmdet

./tools/dist_test configs/sku110k/sku110k_faster_rcnn_r50_fpn_anchor_1x_4tiles_test_half_res.py workdir/faster_rcnn_r50_fpn_anchor_1x_4tiles/latest.pth 2 --format_only --options "jsonfile_prefix=./submit"

7. Convert json output back to SKU110k csv format

python sku110k_scripts/json_out_to_submit.py --args

Experiments

1. Initial experiments

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc mAP AP@0.5 AP@0.75 AR Tr.mAP Tr.AP@0.5 Tr.AP@0.75 Tr.AR
RetinaNet-r50-fpn r50 1x 0.001 2 (1333, 800) 4 (octave) 0.463 0.751 0.532 0.512 0.467 0.752 0.535 0.516
Faster-RCNN-r50-fpn r50 1x 0.005 2 (1333, 800) [8] 0.523 0.850 0.592 0.582 0.537 0.862 0.612 0.594

2. Non-dense anchoring

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc 4tiles mAP AP@0.5 AP@0.75 AR Tr.mAP Tr.AP@0.5 Tr.AP@0.75 Tr.AR
GA-RetinaNet-r50-fpn r50 1x 0.001 2 (816, 1088) 4 (octave) ☐ 0.523 0.870 0.579 0.583 0.532 0.881 0.590 0.591
GA-RetinaNet-x101-32x4d-fpn x101-32x4d 1x 0.001 2 (816, 1088) 4 (octave) ☐ 0.537 0.882 0.602 0.598 0.552 0.896 0.623 0.610
RepPoints-moment-r50-fpn r50 1x 0.02 6 (816, 1088) 4 (base) ☐ 0.505 0.815 0.578 0.562 0.519 0.820 0.601 0.574

3. Comparison of different anchor scales for Faster-RCNN

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc mAP AP@0.5 AP@0.75 AR Tr.mAP Tr.AP@0.5 Tr.AP@0.75 Tr.AR
Faster-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [8] 0.522 0.850 0.591 0.577 0.534 0.862 0.611 0.590
Faster-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [4] 0.551 0.912 0.614 0.613 0.567 0.926 0.636 0.629
Faster-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [3] 0.549 0.911 0.611 0.614

4. Comparison of different anchor scales for RetinaNet

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc mAP AP@0.5 AP@0.75 AR Tr.mAP Tr.AP@0.5 Tr.AP@0.75 Tr.AR
RetinaNet-r50-fpn r50 1x 0.001 2 (1333, 800) 4 (octave) 0.463 0.751 0.532 0.512 0.467 0.752 0.535 0.516
RetinaNet-r50-fpn r50 1x 0.001 2 (1333, 800) 3 (octave) 0.508 0.849 0.564 0.569 0.513 0.853 0.574 0.574

5. Bells and whistles testing

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc 4tiles s-nms test extra augs traintime flip testtime flip mAP AP@0.5 AP@0.75 AR
Faster-RCNN-r50-fpn r50 1x 0.005 2 (752, 1024), (816, 1088), (880, 1152) [4] ☐ ☐ ☐ ✓ ☐ 0.552 0.912 0.615 0.616
Faster-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [4] ☐ ☐ ✓ ☐ ☐ 0.548 0.911 0.608 0.612
Faster-RCNN-r50-fpn r50 2x 0.005 2 (816, 1088) [4] ☐ ☐ ✓ ✓ ☐ 0.540 0.906 0.596 0.606
Faster-RCNN-r50-fpn r50 2x 0.005 2 (816, 1088) [4] ☐ ☐ ✓ ✓ ✓ 0.510 0.888 0.543 0.584

6. Cascade-RCNN comparison

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc 4tiles s-nms test mAP AP@0.5 AP@0.75 AR Tr.mAP Tr.AP@0.5 Tr.AP@0.75 Tr.AR
Cascade-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [8] ☐ ☐ 0.525 0.840 0.604 0.582 0.542 0.862 0.647 0.596
Cascade-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [4] ☐ ☐ 0.553 0.902 0.626 0.615 0.574 0.926 0.653 0.634
Cascade-RCNN-r50-fpn r50 1x 0.005 2 (816, 1088) [4] ☐ ✓ 0.556 0.900 0.632 0.622 0.577 0.925 0.659 0.642
Cascade-RCNN-x101-32x4d-fpn x101-32x4d 1x 0.005 2 (768, 1024) [4] ☐ ☐ 0.556 0.903 0.629 0.617 0.583 0.929 0.665 0.640
Cascade-RCNN-x101-32x4d-fpn x101-32x4d 1x 0.005 2 (768, 1024) [4] ☐ ✓ 0.560 0.902 0.635 0.623 0.585 0.929 0.672 0.647

7. Tiling strategies

Config Backbone Lr schd Base lr imgs_p_gpu img_scale anchor_sc 4tiles s-nms test mAP AP@0.5 AP@0.75 AR
Faster-RCNN-r50-fpn (w/o merging) r50 1x 0.005 2 (816, 1088) [8] ✓ ☐ 0.561 0.912 0.632 0.628
Faster-RCNN-r50-fpn (w/o merging) r50 1x 0.005 2 (816, 1088) [4] ✓ ☐ 0.566 0.928 0.636 0.636
Faster-RCNN-r50-fpn (merged) r50 1x 0.005 2 (816, 1088) [4] ✓ ☐ 0.547 0.894 0.615 0.611
Faster-RCNN-r50-fpn (full frame) r50 1x 0.005 2 (816, 1088) [4] ✓ ✓ 0.577 0.928 0.659 0.654

Citation

Feel free to cite my report if you use any of the results for benchmarking in your work.

@misc{kozlov2020working,
    title={Working with scale: 2nd place solution to Product Detection in Densely Packed Scenes [Technical Report]},
    author={Artem Kozlov},
    year={2020},
    eprint={2006.07825},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}