A wildlife model can miss an animal that is actually present, draw a box around vegetation, or find the right animal and give it the wrong name. These are different failures. A useful result starts by knowing which question the system was meant to answer.
In the ImageSignals video The Lizard The AI Couldn't See, the opening photograph contains a shingleback lizard. The example comes from an Australian field study by Pestell and colleagues. Their Figure 3 records a false negative: MegaDetector missed the lizard. Another panel records the opposite error, with vegetation marked as an animal. The figure also includes undercounting and overcounting examples. [1]
This is a concrete case with a documented source, rather than a picture chosen to make a model look foolish. The errors coexist with strong aggregate detection results in that evaluation. Neither the photograph nor the study establishes the precise cause of that individual miss. Calling it camouflage, a texture shortcut or a lack of reptile training would go beyond the evidence available here.
The more useful lesson is about the chain of decisions between a photograph and a species name. The box and the name need separate evidence.
Finding an animal and naming it are different tasks
Object detection asks where a target is and, often, which broad category it belongs to. Image classification assigns a label to an image or a selected region. A common camera-trap workflow uses a detector to find animals, then gives their crops to a species classifier. It is one workflow, not the architecture of every wildlife system.
MegaDetector's documentation makes its role clear: it locates animals, people and vehicles; it does not identify animal species. Its animal box should therefore not be read as a species identification. A separate classifier can use that region to make a more specific prediction. [2]
Imagine a box correctly surrounding a deer. The box can be useful even if the next stage assigns the wrong deer species. Conversely, a capable species classifier cannot inspect a crop that a detector never produced. In a crop-only route, a missed animal is an upstream failure, before species classification has a chance to help.
That distinction also prevents an unfair conclusion about the opening case. The missed lizard demonstrates a detection error. It does not establish that a species classifier examined the animal and confused its identity.
Why cropping can help—and what it depends on
A distant animal can occupy a small fraction of a large camera-trap photograph. When the entire scene is reduced to a classifier's fixed input size, the animal occupies an even smaller part of that input. Cropping around it first can allocate more of the input to the animal. This preserves more of the available detail through that resizing step; it does not create detail absent from the original.
Gadot and colleagues tested whole-image and crop-based approaches in To crop or not to crop. In Table 6, the unfiltered LILA test condition has overall macro-average F1 of 0.455 for the whole-image classifier and 0.576 for the crop classifier. Those are F1 scores, not percentages of photographs correctly identified. The absolute difference is 0.121 on the F1 scale. [3]
Macro-average F1 gives each evaluated class equal weight in the average. Its interpretation differs from overall accuracy and from an average weighted by class frequency. The result belongs to the paper's label policy, data and evaluation pipeline; it is not a performance estimate for ImageSignals or a promise that cropping always improves recognition.
Cropping changes background as well as the animal's representation. The comparison does not isolate how much of the improvement came from foreground detail and how much came from reducing background cues. It supports the tested crop-based approach, rather than a universal explanation of its gain.
There is still a dependency: a detector-generated crop must exist and contain the relevant animal. It is worth measuring the pipeline from the original files, including misses, rather than reporting only classifications of successful crops. Whole-image and combined routes can be evaluated too; their usefulness must be demonstrated for the chosen task.
A correct box cannot restore missing information
Even a well-selected crop may not contain the evidence needed to distinguish close relatives. A diagnostic marking can be hidden behind a branch, lost to motion blur or unresolved at the available distance. Infrared night images can also lack visible-colour information that a daylight identification would use.
Welbourne and colleagues provide a useful photographic example. Their camera-trap experiments found that two similar Lampropholis skinks could be distinguished with the modified focal distance, while the standard setup did not provide sufficiently sharp distinguishing detail. This was a camera and identification study, not an AI classification trial. [4]
The practical implication is about evidence acquisition: a better frame or better focus may help where another classifier cannot. Enlarging an existing crop makes it easier to inspect, but does not prove that newly apparent texture is an authentic diagnostic feature. A more specific label is justified only when the available evidence and the method support it.
Why 90% accuracy can conceal the failure you care about
Consider a deliberately hypothetical test containing 900 photographs of one common species and 100 of a rare species. A model that predicts the common species every time gets 90% accuracy. Its recall for the rare species is zero. The arithmetic is correct; the headline tells us very little about its value for finding the rare animal.
The example is not a result from the lizard study. It illustrates why the distribution of test labels matters. A conservation project may care most about precisely the species that contributes least to an overall average.
Schneider and colleagues examined class-specific performance, training-set size and new-location testing in a camera-trap study. Their findings show why some underrepresented classes need particular attention, without establishing one minimum image count that guarantees reliable recognition for every species. [5]
Useful reporting includes per-species recall, precision, sample counts and a confusion matrix. Recall asks how many genuine examples were found or correctly labelled. Precision asks how many predictions for a category were correct. The task and denominator must be stated: animal detection and species classification do not share an interchangeable success rate.
The whole pipeline needs its own denominator
Suppose a separate hypothetical collection contains 100 animal images. A detector finds 75 of them. The classifier assigns the correct species to 60 of those 75 crops. Classification accuracy on the detected animals is 80%, but correct species identification across the original 100 images is 60%.
The 25 misses did not disappear from the real task merely because they were absent from the crop set. These are illustrative counts, not measurements or an estimate of either study's results. They show why evaluating each stage separately must be accompanied by an evaluation of the complete workflow.
The reverse dependency matters too. False detections can send vegetation or other non-animal regions into a species classifier. A system needs a way to reject or flag those inputs; a confident species name is not independent confirmation that the detector found an animal.
A familiar camera site is not the whole world
Images from one fixed camera can share backgrounds, viewpoints and lighting. A random image split may put very similar conditions in training and testing. Success there does not settle performance at an unfamiliar site.
Beery, Van Horn and Perona's Recognition in Terra Incognita explicitly tests recognition across camera locations and documents the difficulty of generalizing to new environments. MLWIC2, by Tabak and colleagues, reports both successful transfer and substantial variation across external datasets. New locations do not inevitably defeat every model, but transfer needs to be checked rather than assumed. [6, 7]
For a wildlife deployment, an informative review therefore includes held-out locations and conditions relevant to the actual survey. A result on familiar daytime photographs is not automatically evidence for night images from a different habitat. The labels also need to cover the species or broader categories the system is expected to report.
Choose the level of answer the evidence supports
Sometimes the useful answer is “animal present”, “lizard” or a broader taxonomic group, accompanied by an unresolved species identification. That is a workflow choice requiring its own validation, not a claim that every model can reliably detect when it is uncertain.
Keep the original frame, the selected region and the predicted label available for review. Inspect some images labelled empty as well as positive detections: reviewing only flagged images cannot reveal the animals missed entirely. When a rare-species record matters, another frame, a sequence or expert review may be more valuable than repeatedly asking for a more confident name.
ImageSignals' object detection, region search and image description can support inspection of visible content within their supported capabilities. These general tools are not presented here as validated wildlife species classifiers, and the cited studies did not evaluate ImageSignals. Identifying what a picture depicts is also a different task from deciding whether it was generated by AI.
Return to the lizard. The first question was whether an animal was found. The next would have been what evidence supported its identification. A useful image-classification workflow keeps both questions visible—and measures the cases that never reach the naming stage.
Watch the video
The Lizard The AI Couldn't See · ImageSignals · 4:36.
This article draws on research developed for the ImageSignals video series and expands it for readers. The studies discussed were not independently replicated for this article.
Sources & further reading
- Pestell et al., Smart camera traps and computer vision improve detections of small fauna, Ecosphere, 2025. Figure 3 and the MegaDetector evaluation.
- MegaDetector, official documentation, scope of animal detection versus species classification.
- Gadot et al., To crop or not to crop, IET Computer Vision, 2024. Table 6, unfiltered LILA condition.
- Welbourne et al., Improving Terrestrial Squamate Surveys with Camera-Trap Programming and Hardware Modifications, Animals, 2019. Section 3.2.2 and Figure 7.
- Schneider et al., Three critical factors affecting automated image species recognition performance for camera traps, Ecology and Evolution, 2020.
- Beery, Van Horn & Perona, Recognition in Terra Incognita, ECCV 2018.
- Tabak et al., MLWIC2: Improving accessibility and transferability, Ecology and Evolution, 2020.