Product Detection

The goal is to detect products as quadrilaterals from complex backgrounds. Unitail-Det supports the training and evaluation.
For the each of the two test sets (cross-domain and origin-domain), you need to submit a Pickle file named “cross.pkl” or “origin.pkl”, respectively. The Pickle file is required to contain a dictionary following the format below : Submission Examples
{

"imgs" :

[test_x1.jpg,

test_x2.jpg,

... ...,

test_x3.jpg]

"quad":

[<np.ndarray, float32> (n1, 9), # results from 'test_x1.jpg'

<np.ndarray, float32> (n2, 9),

... ...,

<np.ndarray, float32> (n3, 9)]

}
where each <np.ndarray, float32> contains n quadrilaterals in the order of (x1, y1, x2, y2, x3, y3, x4, y4, score)
The geometric mean of mean average precision (mAP) is calculated on the two testing set as the primary metric.

Text Detection

The goal is to detect text regions from pre-localized product images. Unitail-OCR provides the training and testing. Like ICDAR 2015, no validation set is provided.

Text Recognition

The goal is to recognize words over a set of pre-localized text regions. Unitail-OCR provides the training and testing. Like ICDAR 2015, no validation set is provided.

Product Matching

The goal is to recognize products by matching a set of query samples to the Unitail-OCR gallery. The task is split into two tracks: Hard Example Track, which is evaluated on 2.5k selected hard examples; this track is designed for scenarios in which products are visually similar (for example pharmacy stores); and General Track, which is conducted on all 10k samples.
A pickle file named "matchhard.pkl" or "matchgeneral.pkl" is required for the two tracks, respectively, each should be following the format below : Submission Examples
{

"imgs" :

[test_x1.jpg,

test_x2.jpg,

... ...,

test_x3.jpg]

"cls" :

[c1(int64), # results from 'test_x1.jpg'

c2(int64),

... ...,

c3(int64),

}