The conference paper Jean Ogier Du Terrail, Frédéric Jurie. ON THE USE OF DEEP NEURAL NETWORKS FOR THE DETECTION OF SMALL VEHICLES IN ORTHO-IMAGES. IEEE International Conference on Image Processing, Sep 2017, Beijing, China. (PDF) uses the terms "hard-mining" (6×), "hard mining" (2×), "hard examples" (3×), "hard example mining" (1×), "hard negative" (2×), "hard-negative samples" (1×) and "hard-negative-mining strategies" (1×).

I have no idea what the "hard" specifyer means in this context. As it is mentioned in conjunction with bootstrapping, I suspect that it might be a term from statistics rather than GIS or AI/IR/machine learning/visual object detection or (deep convolutional) artificial neural networks. (It might, of course, be a remote-sensing-specific term.)


2.3. Hard-Mining strategies

Bootstrapping offers a lot of liberties on how the hard examples are chosen. One could for instance pick a limited number of false positives per image or one could fix a threshold and only pick a false positive if its score is superior to a fixed threshold (0.5 for instance). [...]

Does "hard" (in general, or within the terms listed above) mean anything specific in statistics, and if so, what? From the context, I don't suppose that it refers to the difficulty of the problem.

I figured it may be related "hard evidence", but that didn't help me in determining what it might mean here.

  • 8
    $\begingroup$ I think the authors mean that a "hard example" is an example which is poorly-predicted by the model. The example of a score at 0.5 makes sense as an indicator that a particular example is "challenging to classify correctly" in a binary classification problem because it is equidistance between 0 and 1. I'm not aware of any sort of a technical definition -- the usages with which I'm familiar are meant in the colloquial sense of "the model will likely get these things wrong." $\endgroup$
    – Sycorax
    Commented Jul 25, 2017 at 18:19
  • 1
    $\begingroup$ Agree with @Sycorax, just read this in page 3 of the focal loss paper "... by down-weighting the loss of examples with large errors (hard examples)..." $\endgroup$
    – radream
    Commented Jul 5, 2021 at 23:58

1 Answer 1


@Sycorax is correct. "Hard examples" is referring to the examples in the training set that are being mislabeled by the current version of the classifier. Oftentimes it is only used for the background class, which is too large a set for anyone to mine without some kind of a strategy (binary classification on imbalanced sets is hard).

This term was probably coined by Girshick (I think?) in the seminal article DPM and is now widely used in the Object detection community for instance in OHEM, where the negative windows used at each step of the training are chosen according to their current score.

The latter article is an example of Online hard examples mining (hence the title) whereas the ICIP article explores different Offline hard examples mining strategies.

  • 2
    $\begingroup$ "Hard examples" was not coined by Girshick. Mining hard negatives has been particularly employed by Dalal & Triggs in their HOG paper from 2005. $\endgroup$
    – caustic
    Commented May 3, 2020 at 18:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.