Dataset and Protocol

The dataset used in this contest was acquired using a XIMEA snapshot VIS camera. The videos were captured at 25 frames per second (FPS). Each frame was originally captured in 2D with 16 bands arranged in a mosaic mode. Each frame is then converted to 3D with the first two dimensions index the location of each pixel, and the third dimension indexes the band number (code provided). The 16 bands cover the range from 470nm to 620nm, and each band image originally consists of 512×256 pixels. RGB videos were also acquired at the same frame rate in a view point very close to the hyperspectral videos. False-color videos generated from the hyperspectral videos are also provided. The first frames of testing dataset are as follows:

ball basketball board book bus
bus2 campus car car2 car3
card coin coke drive excavator
face face2 forest forest2 fruit
hand kangaroo paper pedestrian pedestrian2
player playground rider1 rider2 rubik
student toy1 toy2 truck worker
Dataset and Source Code Links

Dataset Link: the dataset can be accessed via Google Drive or Baidu YunPan Access code: n616

Evaluation code:

2D image to hyperspectral cube conversion code:

Python code for X2Cube and Evaluation

The car3, basketball sequences in training the car3 sequence in testing are corrected.

Camera Calibration

The camera calibration process involves two steps: dark calibration and spectral correction. Dark calibration aims to remove the influence of noise produced by the camera sensor. It is done by subtracting a dark frame from the captured image, for which the dark frame was captured with lens covered by a cap. The goal of spectral calibration is to reduce the distortion of spectral responses. It is done by applying a sensor-specific spectral correction matrix on the measurement in each pixel. The 16 band images of the corrected hyperspectral data cube are saved as 2D frames with the 16 bands arranged again in a mosaic mode.

Image Registration

The hyperspectral sequences and color sequences are registered to make them describe almost the same scene. This is done by manually selecting matching points in the first frame of both hyperspectral and color videos and then calculating geometrical transformation. The resulted transformation matrix was applied to all subsequent color frames to make alignment with the corresponding hyperspectral frames.

Image Conversion

To ensure fair comparison, the hyperspectral videos were converted to false color videos using CIE color matching functions. This produces strictly spatially aligned hyperspectral and false-color videos.

For details on the above steps, please refer to: F. Xiong, J. Zhou, and Y. Qian. "Material based object tracking in hyperspectral videos", IEEE Trans. Image Process., vol. 29, no. 1, pp. 3719-3733, 2020.


A single upright bounding box is provided for the location of the target object in each frame. The bounding box is represented by the centre location and its height and width. The labels for hyperspectral and color videos were generated independently. The labels for the hyperspectral videos can be used directly on the false-color videos.


The whole dataset contains 40 sets of videos for training and 35 sets of videos for testing. Every video is labelled with associated challenging factors out of eleven attributes, including illumination variation, scale variation, occlusion , deformation, motion blur, fast motion, in-plane rotation, out-of-plane rotation, out-of-view , background clutters, and low resolution.


  • The use of the training set is optional.
  • Tracking starts from the first frame of the sequence. The bounding box in the first frame are used to initialize the location of tracking. Single object tracking is expected.
  • The same model hyper-parameters shall be used for all the sequences.
  • The tracking results contain a sequence of bounding boxes for each frame.

Evaluation Metrics

Precision plot, success plot and area under curve (AUC) will be used to calculate the performance of all the trackers. Precision plot records the fractions of frames whose estimated location is within a given distance threshold to the ground truth. The average distance precision rate is reported at a threshold of 20 pixels. Success plot shows the percentages of successful frames whose overlap ratio between the predicted bounding box and ground-truth is larger than a certain threshold varied from 0 to 1. AUC will be caluclated on each success plot. All the results are presented with one-pass evaluation (OPE), i.e., a tracker is run throughout a test sequence with initialization from the ground truth position in the initial frame. Related codes are provided in the source code package.

Technical Support:

Fengchao Xiong

School of Computer Science and Engineering, Nanjing University Science and Technology