EfficientDet (PyTorch)
A PyTorch implementation of EfficientDet.
It is based on the
- official Tensorflow implementation by Mingxing Tan and the Google Brain team
- paper by Mingxing Tan, Ruoming Pang, Quoc V. Le EfficientDet: Scalable and Efficient Object Detection
There are other PyTorch implementations. Either their approach didn't fit my aim to correctly reproduce the Tensorflow models (but with a PyTorch feel and flexibility) or they cannot come close to replicating MS COCO training from scratch.
Aside from the default model configs, there is a lot of flexibility to facilitate experiments and rapid improvements here -- some options based on the official Tensorflow impl, some of my own:
- BiFPN connections and combination mode are fully configurable and not baked into the model code
- BiFPN and head modules can be switched between depthwise separable or standard convolutions
- Activations, batch norm layers are switchable via arguments (soon config)
- Any backbone in my
timm
model collection that supports feature extraction (features_only
arg) can be used as a bacbkone.
Updates
2020-01-09
- set model config to read-only after creation to reduce likelyhood of misuse
- no accessing model or bench .config attr in forward() call chain (for torcscript compat)
- numerous smaller changes that allow jit scripting of the model or train/predict bench
2020-10-30
Merged a few months of accumulated fixes and additions.
- Proper fine-tuning compatible model init (w/ changeable # classes and proper init, demoed in train.py)
- A new dataset interface with dataset support (via parser classes) for COCO, VOC 2007/2012, and OpenImages V5/Challenge2019
- New focal loss def w/ label smoothing available as an option, support for jit of loss fn for (potential) speedup
- Improved a few hot spots that squeek out a couple % of throughput gains, higher GPU utilization
- Pascal / OpenImages evaluators based on Tensorflow Models Evaluator framework (usable for other datasets as well)
- Support for native PyTorch DDP, SyncBN, and AMP in PyTorch >= 1.6. Still defaults to APEX if installed.
- Non-square input image sizes are allowed for the model (the anchor layout). Specified by image_size tuple in model config. Currently still restricted to
size % 128 = 0
on each dim. - Allow anchor target generation to be done in either dataloader process' via collate or in model as in past. Can help balance compute.
- Filter out unused target cls/box from dataset annotations in fixed size batch tensors before passing to target assigner. Seems to speed convergence.
- Letterbox aware Random Erasing augmentation added.
- A (very slow) SoftNMS impl added for inference/validation use. It can be manually enabled right now, can add arg if demand.
- Tested with PyTorch 1.7
- Add ResDet50 model weights, 41.6 mAP.
A few things on priority list I haven't tackled yet:
- Mosaic augmentation
- bbox IOU loss (tried a bit but so far not a great result, need time to debug/improve)
NOTE There are some breaking changes:
- Predict and Train benches now output XYXY boxes, NOT XYWH as before. This was done to support other datasets as XYWH is COCO's evaluator requirement.
- The TF Models Evaluator operates on YXYX boxes like the models. Conversion from XYXY is currently done by default. Why don't I just keep everything YXYX? Because PyTorch GPU NMS operates in XYXY.
- You must update your version of
timm
to the latest (>=0.3), as some APIs for helpers changed a bit.
Training sanity checks were done on VOC and OI
- 80.0 @ 50 mAP finetune on voc0712 with no attempt to tune params (roughly as per command below)
- 18.0 mAP @ 50 for OI Challenge2019 after couple days of training (only 6 epochs, eek!). It's much bigger, and takes a LOONG time, many classes are quite challenging.
2020-09-03
- All models updated to latest checkpoints from TF original.
- Add experimental soft-nms code, must be manually enabled right now. It is REALLY slow, .1-.2 mAP increase.
2020-07-27
- Add updated TF ported weights for D3 model (better training) and model def and weights for new D7X model (54.3 val mAP)
- Fix Windows bug so it at least trains in non-distributed mode
2020-06-15
Add updated D7 weights from Tensorflow impl, 53.1 validation mAP here (53.4 in TF)
2020-06-14
New model results, I've trained a D1 model with some WIP augmentation enhancements (not commited), just squeaking by official weights.
EfficientDet-D1:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.393798
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.586831
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.420305
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.191880
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.455586
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.571316
Also, Soyeb Nagori trained an EfficientDet-Lite0 config using this code and contributed the weights.
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.319861
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.500062
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.336777
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.111257
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.378062
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.501938
Unlike the other tf_ prefixed models this is not ported from (as of yet unreleased) TF official model, but it used
TF ported weights from timm
for the pretrained imagenet model as the backbone init, thus it uses SAME padding.
Models
The table below contains models with pretrained weights. There are quite a number of other models that I have defined in model configurations that use various timm
backbones.
Variant | mAP (val2017) | mAP (test-dev2017) | mAP (TF official val2017) | mAP (TF official test-dev2017) |
---|---|---|---|---|
tf_efficientdet_lite0.pth | 32.0 | TBD | N/A | N/A |
efficientdet_d0.pth | 33.6 | TBD | 33.5 | 33.8 |
tf_efficientdet_d0.pth | 34.2 | TBD | 34.3 | 34.6 |
efficientdet_d1.pth | 39.4 | 39.5 | 39.1 | 39.6 |
tf_efficientdet_d1.pth | 40.1 | TBD | 40.2 | 40.5 |
resdet50 | 41.6 | TBD | N/A | N/A |
tf_efficientdet_d2.pth | 43.4 | TBD | 42.5 | 43 |
tf_efficientdet_d3.pth | 47.1 | TBD | 47.2 | 47.5 |
tf_efficientdet_d4.pth | 49.2 | TBD | 49.3 | 49.7 |
tf_efficientdet_d5.pth | 51.2 | TBD | 51.2 | 51.5 |
tf_efficientdet_d6.pth | 52.0 | TBD | 52.1 | 52.6 |
tf_efficientdet_d7.pth | 53.1 | 53.4 | 53.4 | 53.7 |
tf_efficientdet_d7x.pth | 54.3 | TBD | 54.4 | 55.1 |
See model configurations for model checkpoint urls and differences.
NOTE: Official scores for all modules now using soft-nms, but still using normal NMS here.
NOTE: In training some experimental models, I've noticed some potential issues with the combination of synchronized BatchNorm (--sync-bn
) and model EMA weight everaging (--model-ema
) during distributed training. The result is either a model that fails to converge, or appears to converge (training loss) but the eval loss (running BN stats) is garbage. I haven't observed this with EfficientNets, but have with some backbones like CspResNeXt, VoVNet, etc. Disabling either EMA or sync bn seems to eliminate the problem and result in good models. I have not fully characterized this issue.
Environment Setup
Tested in a Python 3.7 or 3.8 conda environment in Linux with:
- PyTorch 1.4, 1.6, or 1.7 (I recommend avoiding PyTorch 1.5 due to some jit and argmax issues)
- PyTorch Image Models (timm) >= 0.3.0,
pip install timm
or local install from (https://github.com/rwightman/pytorch-image-models) - Apex AMP master (as of 2020-08)
NOTE - There is a conflict/bug with Numpy 1.18+ and pycocotools 2.0, force install numpy <= 1.17.5 or ensure you install pycocotools >= 2.0.2
Dataset Setup and Use
COCO
MSCOCO 2017 validation data:
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip val2017.zip
unzip annotations_trainval2017.zip
MSCOCO 2017 test-dev data:
wget http://images.cocodataset.org/zips/test2017.zip
unzip -q test2017.zip
wget http://images.cocodataset.org/annotations/image_info_test2017.zip
unzip image_info_test2017.zip
COCO Evaluation
Run validation (val2017 by default) with D2 model: python validate.py /localtion/of/mscoco/ --model tf_efficientdet_d2
Run test-dev2017: python validate.py /localtion/of/mscoco/ --model tf_efficientdet_d2 --split testdev
COCO Training
./distributed_train.sh 4 /mscoco --model tf_efficientdet_d0 -b 16 --amp --lr .09 --warmup-epochs 5 --sync-bn --opt fusedmomentum --model-ema
NOTE:
- Training script currently defaults to a model that does NOT have redundant conv + BN bias layers like the official models, set correct flag when validating.
- I've only trained with img mean (
--fill-color mean
) as the background for crop/scale/aspect fill, the official repo uses black pixel (0) (--fill-color 0
). Both likely work fine. - The official training code uses EMA weight averaging by default, it's not clear there is a point in doing this with the cosine LR schedule, I find the non-EMA weights end up better than EMA in the last 10-20% of training epochs
- The default h-params is a very close to unstable (exploding loss), don't try using Nesterov momentum. Try to keep the batch size up, use sync-bn.
Pascal VOC
2007, 2012, and combined 2007 + 2012 w/ labeled 2007 test for validation are supported.
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
find . -name '*.tar' -exec tar xf {} \;
There should be a VOC2007
and VOC2012
folder within VOCdevkit
, dataset root for cmd line will be VOCdevkit.
Alternative download links, slower but up more often than ox.ac.uk:
http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
http://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar
VOC Evaluation
Evaluate on VOC2012 validation set:
python validate.py /data/VOCdevkit --model efficientdet_d0 --num-gpu 2 --dataset voc2007 --checkpoint mychekpoint.pth --num-classes 20
VOC Training
Fine tune COCO pretrained weights to VOC 2007 + 2012:
/distributed_train.sh 4 /data/VOCdevkit --model efficientdet_d0 --dataset voc0712 -b 16 --amp --lr .008 --sync-bn --opt fusedmomentum --warmup-epochs 3 --model-ema --model-ema-decay 0.9966 --epochs 150 --num-classes 20 --pretrained
OpenImages
Setting up OpenImages dataset is a commitment. I've tried to make it a bit easier wrt to the annotations, but grabbing the dataset is still going to take some time. It will take approx 560GB of storage space.
To download the image data, I prefer the CVDF packaging. The main OpenImages dataset page, annotations, dataset license info can be found at: https://storage.googleapis.com/openimages/web/index.html
CVDF Images Download
Follow the s3 download directions here: https://github.com/cvdfoundation/open-images-dataset#download-images-with-bounding-boxes-annotations
Each train_<x>.tar.gz
should be extracted to train/<x>
folder, where x is a hex digit from 0-F. validation.tar.gz
can be extracted as flat files into validation/
.
Annotations Download
Annotations can be downloaded separately from the OpenImages home page above. For convenience, I've packaged them all together with some additional 'info' csv files that contain ids and stats for all image files. My datasets rely on the <set>-info.csv
files. Please see https://storage.googleapis.com/openimages/web/factsfigures.html for the License of these annotations. The annotations are licensed by Google LLC under CC BY 4.0 license. The images are listed as having a CC BY 2.0 license.
wget https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1-anno/openimages-annotations.tar.bz2
wget https://github.com/rwightman/efficientdet-pytorch/releases/download/v0.1-anno/openimages-annotations-challenge-2019.tar.bz2
find . -name '*.tar.bz2' -exec tar xf {} \;
Layout
Once everything is downloaded and extracted the root of your openimages data folder should contain:
annotations/<csv anno for openimages v5/v6>
annotations/challenge-2019/<csv anno for challenge2019>
train/0/<all the image files starting with '0'>
.
.
.
train/f/<all the image files starting with 'f'>
validation/<all the image files in same folder>
OpenImages Training
Training with Challenge2019 annotations (500 classes):
./distributed_train.sh 4 /data/openimages --model efficientdet_d0 --dataset openimages-challenge2019 -b 7 --amp --lr .042 --sync-bn --opt fusedmomentum --warmup-epochs 1 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.999966 --epochs 100 --remode pixel --reprob 0.15 --recount 4 --num-classes 500 --val-skip 2
The 500 (Challenge2019) or 601 (V5/V6) class head for OI takes up a LOT more GPU memory vs COCO. You'll likely need to half batch sizes.
Examples of Training / Fine-Tuning on Custom Datasets
The models here have been used with custom training routines and datasets with great results. There are lots of details to figure out so please don't file any 'I get crap results on my custom dataset issues'. If you can illustrate a reproducible problem on a public, non-proprietary, downloadable dataset, with public github fork of this repo including working dataset/parser implementations, I MAY have time to take a look.
Examples:
- Alex Shonenkov has a clear and concise Kaggle kernel which illustrates fine-tuning these models for detecting wheat heads: https://www.kaggle.com/shonenkov/training-efficientdet (this is out of date wrt to latest versions here)
If you have a good example script or kernel training these models with a different dataset, feel free to notify me for inclusion here...
Results
My Training
EfficientDet-D0
Latest training run with .336 for D0 (on 4x 1080ti):
./distributed_train.sh 4 /mscoco --model efficientdet_d0 -b 22 --amp --lr .12 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.9999
These hparams above resulted in a good model, a few points:
- the mAP peaked very early (epoch 200 of 300) and then appeared to overfit, so likely still room for improvement
- I enabled my experimental LR noise which tends to work well with EMA enabled
- the effective LR is a bit higher than official. Official is .08 for batch 64, this works out to .0872
- drop_path (aka survival_prob / drop_connect) rate of 0.1, which is higher than the suggested 0.0 for D0 in official, but lower than the 0.2 for the other models
- longer EMA period than default
VAL2017
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.336251
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.521584
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.356439
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.123988
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.395033
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.521695
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.287121
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.441450
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.467914
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.197697
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.552515
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.689297
EfficientDet-D1
Latest run with .394 mAP (on 4x 1080ti):
./distributed_train.sh 4 /mscoco --model efficientdet_d1 -b 10 --amp --lr .06 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.99995
For this run I used some improved augmentations, still experimenting so not ready for release, should work well without them but will likely start overfitting a bit sooner and possibly end up a in the .385-.39 range.
Ported Tensorflow weights
TEST-DEV2017
NOTE: I've only tried submitting D7 to dev server for sanity check so far
TF-EfficientDet-D7
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.534
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.726
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.577
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.356
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.569
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.660
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.397
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.644
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.682
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.508
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.718
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.818
VAL2017
TF-EfficientDet-D0
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.341877
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.525112
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.360218
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.131366
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.399686
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.537368
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.293137
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.447829
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.472954
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.195282
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.558127
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.695312
TF-EfficientDet-D1
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.401070
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.590625
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.422998
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.211116
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.459650
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.577114
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.326565
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.507095
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.537278
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.308963
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.610450
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.731814
TF-EfficientDet-D2
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.434042
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.627834
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.463488
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.237414
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.486118
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.606151
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.343016
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.538328
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.571489
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.350301
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.638884
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.746671
TF EfficientDet-D3
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.471223
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.661550
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.505127
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.301385
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.518339
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.626571
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.365186
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.582691
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.617252
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.424689
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.670761
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.779611
TF-EfficientDet-D4
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.491759
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.686005
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.527791
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.325658
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.536508
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.635309
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.373752
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.601733
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.638343
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.463057
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.685103
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.789180
TF-EfficientDet-D5
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.511767
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.704835
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.552920
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.355680
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.551341
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.650184
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.384516
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.619196
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.657445
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.499319
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.695617
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.788889
TF-EfficientDet-D6
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.520200
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.713204
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.560973
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.361596
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.567414
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.657173
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.387733
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.629269
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.667495
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.499002
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.711909
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.802336
TF-EfficientDet-D7
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.531256
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.724700
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.571787
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.368872
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.573938
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.668253
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.393620
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.637601
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.676987
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.524850
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.717553
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.806352
TF-EfficientDet-D7X
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.543
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.737
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.585
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.401
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.579
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.680
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.398
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.649
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.689
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.550
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.725
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.823
TODO
- Basic Training (object detection) reimplementation
- Mosaic Augmentation
- Rand/AutoAugment
- BBOX IoU loss (giou, diou, ciou, etc)
- Training (semantic segmentation) experiments
- Integration with Detectron2 / MMDetection codebases
- Addition and cleanup of EfficientNet based U-Net and DeepLab segmentation models that I've used in past projects
- Addition and cleanup of OpenImages dataset/training support from a past project
- Exploration of instance segmentation possibilities...
If you are an organization is interested in sponsoring and any of this work, or prioritization of the possible future directions interests you, feel free to contact me (issue, LinkedIn, Twitter, hello at rwightman dot com). I will setup a github sponser if there is any interest.