Spaces:
Sleeping
app_file: app.py
colorFrom: yellow
colorTo: green
description: 'TODO: add a description here'
emoji: 🤑
pinned: false
runme:
id: 01HPS3ASFJXVQR88985QNSXVN1
version: v3
sdk: gradio
sdk_version: 4.36.0
tags:
- evaluate
- metric
title: user-friendly-metrics
How to Use
import evaluate
from seametrics.payload.processor import PayloadProcessor
payload = PayloadProcessor(
dataset_name="SENTRY_VIDEOS_DATASET_QA",
gt_field="ground_truth_det_fused_id",
models=["ahoy_IR_b2_engine_3_7_0_757_g8765b007_oversea"],
sequence_list=["Sentry_2023_02_08_PROACT_CELADON_@6m_MOB_2023_02_08_14_41_51"],
# tags=["GT_ID_FUSION"],
tracking_mode=True
).payload
module = evaluate.load("SEA-AI/user-friendly-metrics")
res = module._compute(payload, max_iou=0.5, recognition_thresholds=[0.3, 0.5, 0.8])
print(res)
{
"global": {
"ahoy_IR_b2_engine_3_6_0_49_gd81d3b63_oversea": {
"all": {
"f1": 0.15967351103175614,
"fn": 2923.0,
"fp": 3666.0,
"num_gt_ids": 10,
"precision": 0.14585274930102515,
"recall": 0.1763877148492533,
"recognition_0.3": 0.1,
"recognition_0.5": 0.1,
"recognition_0.8": 0.1,
"recognized_0.3": 1,
"recognized_0.5": 1,
"recognized_0.8": 1,
"tp": 626.0
}
}
},
"per_sequence": {
"Sentry_2023_02_08_PROACT_CELADON_@6m_MOB_2023_02_08_12_51_49": {
"ahoy_IR_b2_engine_3_6_0_49_gd81d3b63_oversea": {
"all": {
"f1": 0.15967351103175614,
"fn": 2923.0,
"fp": 3666.0,
"num_gt_ids": 10,
"precision": 0.14585274930102515,
"recall": 0.1763877148492533,
"recognition_0.3": 0.1,
"recognition_0.5": 0.1,
"recognition_0.8": 0.1,
"recognized_0.3": 1,
"recognized_0.5": 1,
"recognized_0.8": 1,
"tp": 626.0
}
}
}
}
}
Metric Settings
The max_iou parameter is used to filter out the bounding boxes with IOU less than the threshold. The default value is 0.5. This means that if a ground truth and a predicted bounding boxes IoU value is less than 0.5, then the predicted bounding box is not considered for association. So, the higher the max_iou value, the more the predicted bounding boxes are considered for association.
Output
The output is a dictionary containing the following metrics:
| Name | Description |
|---|---|
| recall | Number of detections over number of objects. |
| precision | Number of detected objects over sum of detected and false positives. |
| f1 | F1 score |
| num_gt_ids | Number of unique objects on the ground truth |
| fn | Number of false negatives |
| fp | Number of of false postives |
| tp | number of true positives |
| recognized_th | Total number of unique objects on the ground truth that were seen more then th% of the times |
| recognition_th | Total number of unique objects on the ground truth that were seen more then th% of the times over the number of unique objects on the ground truth |
How it Works
We levereage one of the internal variables of motmetrics MOTAccumulator class, events, which keeps track of the detections hits and misses. These values are then processed via the track_ratios function which counts the ratio of assigned to total appearance count per unique object id. We then define the recognition function that counts how many objects have been seen more times then the desired threshold.
Citations
@InProceedings{huggingface:module,
title = {A great new module},
authors={huggingface, Inc.},
year={2020}}
@article{milan2016mot16,
title={MOT16: A benchmark for multi-object tracking},
author={Milan, Anton and Leal-Taix{\'e}, Laura and Reid, Ian and Roth, Stefan and Schindler, Konrad},
journal={arXiv preprint arXiv:1603.00831},
year={2016}}