Skip to content

Latest commit

 

History

History
60 lines (43 loc) · 3.7 KB

model_zoo.md

File metadata and controls

60 lines (43 loc) · 3.7 KB
Raw

Model Zoos

We provide the following pre-trained SiamMOT models.

The name follows the convention SiamMOT-BACKBONE-MOTION_MODEL

Airbone Object Tracking Challenge (AOT)

You can find the details of the challenge in the offical homepage hosted by AICrowd. Top methods will share the cash prize pool of $50,000.

The baseline results in the AOT leaderboard are provided with this model: SiamMOT-DLA34-EMM

MOTChallenge-2017 Test (Public detection)

In order to run the model with the public detection, you need to

  1. set INFERENCE.USE_GIVEN_DETECTION = True in the configuration file

  2. set INPUT.AMODAL = True in the configuration file (MOT17 uses amodal bounding box annotations)

  3. ingest the public detection to DataSample object, we provide the ingested public detection. After extraction, they should be placed under dataset_root/annotation folder.

Model Training data MOTA IDF1
SiamMOT-DLA34-EMM CrowdHuman 65.01 61.86
SiamMOT-DLA34-EMM CrowdHuman + MOT17 66.09 63.49
SiamMOT-DLA34-IMM CrowdHuman - -
SiamMOT-DLA34-IMM CrowdHuman + MOT17 - -

TAO-Person

Use the default configuration file to generate the following results.

Model Training data TAP@0.5 TAP@0.75
SiamMOT-DLA34-EMM CrowdHuman + COCO17 37.98 19.99
SiamMOT-DLA169-EMM CrowdHuman + COCO17 - -

Pre-trained Faster-RCNN on COCO-2017

The following models (Faster-RCNN with FPN) can be used to initialize SiamMOT, and they are pre-trained on COCO-2017 80-class object detection dataset (training split). The following table summarize their results in COCO-2017 80-class validation set.

In order to initiate SiamMOT during training, download the corresponding model weight, and point its path to MODEL.WEIGHT in the configuration file.

Backbone box-MAP
DLA-34 35.9
DLA-102 38.3
DLA-169 39.8
DLA-102-DCN 42.0
DLA-169-DCN 42.9
ResNet-50 37.3
ResNet-101 39.5
ResNet-50-DCN 40.6
ResNet-101-DCN 43.0