Top 30 Easy AI & ML Interview Questions and Answers for Freshers

30 questions. Pick an answer, or open “Show the answer” to read it.

Answers are graded in your browser. Quiz progress is not saved, and no XP or streak is earned here. The app keeps score.

1. Question 1 of 30

   When converting a large list of flat dictionaries to a pandas DataFrame, which method is the idiomatic one-shot approach for aligning keys and avoiding repeated DataFrame growth?

   • aConvert the list of dicts into a list of lists first, then pass that to pd.DataFrame
   • bUse DataFrame.from_dict with orient='index' and then transpose the result
   • cPass the list of dicts directly to pd.DataFrame so the constructor aligns the dictionary keys in one operation
   • dIterate through the list and build the DataFrame incrementally using pd.concat for each dictionary
   Show the answer

   Answer: c · Pass the list of dicts directly to pd.DataFrame so the constructor aligns the dictionary keys in one operation

   Passing the list directly to pd.DataFrame is the idiomatic approach for flat row-oriented dictionaries: it aligns the union of keys and fills missing values with NaN. Repeatedly growing a DataFrame can cause unnecessary copying, while the exact performance of alternatives depends on the data and pandas version.

   Read the full bite: Most efficient way to convert list of dicts to pandas DataFrame

2. Question 2 of 30

   For a DataFrame with a string index, which call raises an error and why?

   • adf.loc[:, :], because full slices are disallowed
   • bdf.iloc[0], because position 0 is reserved
   • cdf.loc['x'], because loc cannot take a single label
   • ddf.iloc['x'], because iloc requires integer positions not labels
   Show the answer

   Answer: d · df.iloc['x'], because iloc requires integer positions not labels

   iloc is strictly positional and rejects a string label, raising a TypeError. df.loc['x'] and df.iloc[0] both work correctly on a string-indexed frame.

   Read the full bite: Pandas loc versus iloc indexing

3. Question 3 of 30

   A pandas DataFrame has 5% missing values in customer_age, completely at random. When is dropna the better choice over fillna?

   • aWhen building a churn model where age is predictive and each row contains a labeled outcome
   • bWhen the column uses nullable Int64 dtype to avoid coercion to float64
   • cWhen generating a quick aggregate report and the remaining 95% of data is sufficient
   • dWhen you want to add an is_age_missing indicator column to preserve missingness signal
   Show the answer

   Answer: c · When generating a quick aggregate report and the remaining 95% of data is sufficient

   The card states that dropna is safe for a quick aggregate report when data is missing completely at random and the remaining sample is still large enough. Option A describes the opposite scenario: when age is predictive, dropping rows discards valuable labeled outcomes and reduces model performance, so imputation is preferred instead.

   Read the full bite: Describe strategies for handling missing values in pandas DataFrames

4. Question 4 of 30

   Which query pattern correctly identifies customers who have never placed an order?

   • aFilter with WHERE customer_id NOT IN (SELECT customer_id FROM orders)
   • bLEFT JOIN orders then filter WHERE orders.customer_id IS NULL
   • cLEFT JOIN orders then filter WHERE orders.customer_id = NULL
   • dINNER JOIN customers to orders and select DISTINCT names
   Show the answer

   Answer: b · LEFT JOIN orders then filter WHERE orders.customer_id IS NULL

   B uses the correct anti-join pattern: LEFT JOIN preserves all customers and IS NULL keeps only those with no matching orders. D is not an anti-join, and A can be unsafe when the subquery contains NULL customer IDs because SQL three-valued logic can suppress expected rows.

   Read the full bite: Find customers who have not placed any orders

5. Question 5 of 30

   Which approach correctly handles the full response lifecycle when fetching JSON from a REST API in production using Python's requests?

   • aSet a request timeout but skip status checks to parse faster
   • bUse r.text and manually split the string to extract JSON fields
   • cVerify the status with r.raise_for_status() before parsing with r.json()
   • dPass query parameters in the request body and call r.json() directly
   Show the answer

   Answer: c · Verify the status with r.raise_for_status() before parsing with r.json()

   Calling r.raise_for_status() before trusting r.json() raises an HTTPError for unsuccessful responses before application code treats their bodies as successful payloads. r.json() parses the response body independently and can also raise a JSON decoding error when the body is not valid JSON; timeouts and connection errors need separate handling.

   Read the full bite: How do you fetch JSON from a REST API and parse it?

6. Question 6 of 30

   Why is a violinplot generally preferred over a bar plot when exploring house prices across neighborhoods?

   • aIt preserves the full distribution, revealing features like bimodality and outliers that a bar plot's single statistic would obscure.
   • bIt is the default choice for non-technical audiences because it requires no explanation of medians or quartiles.
   • cIt avoids overplotting by displaying every individual data point in a beeswarm arrangement.
   • dIt summarizes each neighborhood with a single estimate and confidence interval for faster interpretation.
   Show the answer

   Answer: a · It preserves the full distribution, revealing features like bimodality and outliers that a bar plot's single statistic would obscure.

   The card states that bar plots collapse a continuous distribution into a single estimate like the mean, hiding variance, multimodality, and extreme values, whereas violinplots reveal the full distribution shape. Option D actually describes the key limitation of a bar plot, not an advantage of a violinplot.

   Read the full bite: Which plot visualizes a continuous versus categorical variable and why?

7. Question 7 of 30

   During EDA, a heatmap reveals a strong positive correlation between daily ice cream sales and sunscreen sales. What is the most appropriate next step before making causal business recommendations?

   • aDismiss the correlation as irrelevant because correlation metrics cannot prove causation and therefore provide no actionable insight.
   • bLaunch a joint promotion because the heatmap indicates a causal link between the two product categories.
   • cImmediately remove one variable from the dataset since two highly correlated features are always redundant in predictive models.
   • dInvestigate potential confounders such as outdoor temperature and examine partial correlations before making causal recommendations.
   Show the answer

   Answer: d · Investigate potential confounders such as outdoor temperature and examine partial correlations before making causal recommendations.

   The card emphasizes that a strong heatmap correlation should prompt a search for confounders such as temperature and further analysis like partial correlation, not a causal conclusion. Option B is wrong because heatmap color intensity reflects linear association, not causal strength, which is a key red flag mentioned.

   Read the full bite: Explain correlation vs causation and heatmap limitations in EDA

8. Question 8 of 30

   A data scientist chooses Z-score standardization over Min-Max scaling before training a logistic regression model with gradient descent. What is the primary statistical reason for this choice?

   • aMin-Max scaling bounds every feature to the same range, which prevents any single feature from dominating gradient descent weight updates.
   • bZ-score standardization centers features at mean zero with unit variance, ensuring weight updates proceed at comparable rates during gradient descent.
   • cZ-score standardization eliminates outliers entirely, producing a cleaner loss surface than Min-Max scaling would.
   • dLogistic regression requires no scaling because it relies on rank ordering and is therefore scale-invariant like tree-based models.
   Show the answer

   Answer: b · Z-score standardization centers features at mean zero with unit variance, ensuring weight updates proceed at comparable rates during gradient descent.

   Z-score standardization places features on comparable scales so gradient descent updates all weights at similar rates and no feature dominates the loss surface. The Min-Max distractor is tempting but incorrect because bounding values to [0, 1] does not equalize variances, and high-variance features can still compress the signal of others.

   Read the full bite: Min-Max scaling vs Z-score standardization: differences and algorithm preferences

9. Question 9 of 30

   When a categorical feature has thousands of unique values, which approach avoids the dimensionality and memory issues of one-hot encoding while keeping predictive signal?

   • aLabel encoding, which assigns a unique integer to each category so the feature remains a single column
   • bOrdinal encoding, which sorts categories by frequency and replaces them with their rank
   • cDropping the feature to reduce model complexity and training time
   • dTarget encoding, which replaces each category with the mean of the target variable for that category
   Show the answer

   Answer: d · Target encoding, which replaces each category with the mean of the target variable for that category

   Target encoding collapses a high-cardinality categorical feature into a numeric statistic such as a target mean, avoiding one sparse column per category. Compute the statistic out-of-fold for training data and from training data only for validation or test data, with smoothing where appropriate, to reduce target leakage. Label and ordinal encoding can impose a false order on nominal categories.

   Read the full bite: Why avoid one-hot encoding for high cardinality and what are alternatives?

10. Question 10 of 30

    In fraud detection with 1% fraud prevalence, a model labels every transaction as legitimate and achieves 99% accuracy. What is the key lesson?

    • aAccuracy can be misleading because it ignores the rare positive class entirely
    • bMaximizing recall alone is the best strategy since catching fraud matters most
    • cThe model should be accepted because 99% accuracy exceeds most business benchmarks
    • dHigh accuracy guarantees that a model has strong predictive power
    Show the answer

    Answer: a · Accuracy can be misleading because it ignores the rare positive class entirely

    The card describes the accuracy paradox, where a trivial classifier achieves 99% accuracy on imbalanced fraud data while catching zero actual fraud, proving accuracy can hide complete failure on the minority class. Option C is tempting because 99% sounds excellent, but it wrongly rewards a model that is entirely useless for the business goal of detecting fraud.

    Read the full bite: Why is accuracy misleading for fraud detection, and what metrics instead?

11. Question 11 of 30

    Which description best captures how K-Means iteratively forms its final clusters?

    • aIt minimizes straight-line Euclidean distance by iteratively shifting each centroid to the median of its assigned points.
    • bIt selects K random points as fixed centers and assigns every other point to the nearest center in one step.
    • cIt divides the data into K equally sized partitions and then computes the mean of each partition.
    • dIt initializes K centroids, then repeatedly assigns points to the nearest centroid and recomputes each centroid as the cluster mean until stable.
    Show the answer

    Answer: d · It initializes K centroids, then repeatedly assigns points to the nearest centroid and recomputes each centroid as the cluster mean until stable.

    K-Means uses Lloyd's algorithm to repeatedly assign observations to the nearest centroid and update each centroid to the cluster mean until convergence. Option B is tempting because random initialization is correct, but the algorithm is not a single-pass assignment; without iterative refinement it would not minimize within-cluster variance.

    Read the full bite: How does K-Means clustering work and how do you choose K?

12. Question 12 of 30

    You run an experiment and obtain p = 0.03. Which statement correctly interprets this p-value?

    • aThere is a 97% probability that the alternative hypothesis is true.
    • bThere is a 3% chance that the null hypothesis is false.
    • cIf the null hypothesis were true, there is a 3% probability of observing data at least this extreme.
    • dThe result has only a 3% probability of being due to random chance.
    Show the answer

    Answer: c · If the null hypothesis were true, there is a 3% probability of observing data at least this extreme.

    A p-value assumes the null hypothesis is true and quantifies the probability of seeing data at least this extreme under that assumption; it does not measure the probability that the null is false, that the alternative is true, or that the result occurred by chance.

    Read the full bite: What is a p-value? Interpret p = 0.03 at alpha = 0.05.

13. Question 13 of 30

    Which hypothesis pair best tests whether a green signup button increases registrations compared to blue?

    • aH0: There is no difference in registration conversion rate; H1: The registration conversion rates are not equal.
    • bH0: The registration conversion rate for green equals that for blue; H1: The registration conversion rate for green is strictly greater than that for blue.
    • cH0: The blue button has a higher registration conversion rate than the green button; H1: The green button has a higher registration conversion rate than the blue button.
    • dH0: The green and blue buttons perform the same; H1: The green button performs better.
    Show the answer

    Answer: b · H0: The registration conversion rate for green equals that for blue; H1: The registration conversion rate for green is strictly greater than that for blue.

    The correct answer frames H0 as no difference and H1 as a directional increase in the specific metric, matching the one-tailed nature of the business question. Option A is tempting because it uses 'no difference,' but it wrongly uses a two-tailed alternative that ignores the directional ask and wastes statistical power.

    Read the full bite: What are your null and alternative hypotheses for this A/B test?

14. Question 14 of 30

    Which approach best demonstrates a sound framework for measuring e-commerce user engagement?

    • aUse DAU/MAU, adoption, retention, and stickiness tied to conversion and churn
    • bTrack total page views and raw session counts as primary success indicators
    • cCollect qualitative feedback as the main quantifiable engagement signal
    • dMonitor as many metrics as possible without ranking their business impact
    Show the answer

    Answer: a · Use DAU/MAU, adoption, retention, and stickiness tied to conversion and churn

    Among these choices, A is the best framework because it connects engagement measures to business outcomes instead of treating page views as success. For an e-commerce product, apply that principle to the journey: product views and search, add-to-cart, checkout, purchase, repeat purchase, conversion, revenue or margin, and retention. DAU/MAU can remain a supporting habit signal.

    Read the full bite: What metrics track e-commerce user engagement and how do you prioritize them?

15. Question 15 of 30

    Why does replacing sigmoid with ReLU in a deep network reduce vanishing gradients?

    • aReLU's output is bounded between 0 and 1, stabilizing the gradient
    • bReLU normalizes activations to zero mean and unit variance each layer
    • cReLU has a derivative of 1 for positive inputs, so it does not shrink gradients there
    • dReLU multiplies gradients by a factor greater than 1, amplifying them
    Show the answer

    Answer: c · ReLU has a derivative of 1 for positive inputs, so it does not shrink gradients there

    ReLU's unit derivative in its active region avoids the sub-one multiplications that cause sigmoid's gradients to vanish across depth. ReLU is unbounded above, does not normalize, and does not amplify gradients beyond one.

    Read the full bite: Vanishing Gradients and Why ReLU Helps

16. Question 16 of 30

    Why is HSV often preferred over RGB for segmenting a colored object across a scene with mixed sunlight and shadow?

    • aHSV can represent colors that RGB cannot display
    • bHue stays relatively stable under lighting changes, so color separates from brightness
    • cHSV uses fewer bits and processes faster
    • dRGB cannot represent the color red accurately
    Show the answer

    Answer: b · Hue stays relatively stable under lighting changes, so color separates from brightness

    HSV isolates hue from value, so a color keeps its hue under varying illumination, easing thresholding. Both spaces cover the same color gamut, so the capacity claim is false.

    Read the full bite: RGB versus HSV color spaces

17. Question 17 of 30

    How should a healthy production ML lifecycle be structured from start to finish?

    • aAs a linear waterfall from algorithm selection to deployment with no iteration
    • bAs a process that starts with data ingestion and ends once the model is packaged for release
    • cAs an iterative loop starting with business problem framing and continuing through post-deployment monitoring
    • dAs a cycle focused primarily on training models and optimizing offline accuracy metrics
    Show the answer

    Answer: c · As an iterative loop starting with business problem framing and continuing through post-deployment monitoring

    The correct answer is B because the card describes the lifecycle as an end-to-end engineering process that begins with business goal definition and requires continuous monitoring and feedback loops after deployment. The most tempting distractor is D because treating packaging as the final step omits critical monitoring, retraining, and validation stages that keep a production model healthy.

    Read the full bite: Describe the key stages of a typical ML lifecycle

18. Question 18 of 30

    Which property of one-hot vectors prevents a model from inferring that 'excellent' and 'outstanding' are similar?

    • aThey are mutually orthogonal, giving every pair zero cosine similarity
    • bThey use binary values instead of continuous real numbers
    • cThey must be paired with manually crafted synonym lists
    • dThey are too high-dimensional to store in memory efficiently
    Show the answer

    Answer: a · They are mutually orthogonal, giving every pair zero cosine similarity

    A is correct: one-hot vectors are mutually orthogonal, so their cosine similarity is zero and they provide no geometric signal that related words are similar. High dimensionality is a separate practical drawback.

    Read the full bite: Explain word embeddings and why they beat one-hot encoding for large vocabularies

19. Question 19 of 30

    An e-commerce site sees search users convert at 5% and non-search users at 2%. What is the best next step to estimate the true incremental impact of search?

    • aMatch searchers to similar non-searchers using device, landing page, and prior sessions, then compare rates
    • bRun a two-sample t-test on the conversion rates to confirm the difference is significant
    • cImmediately design an A/B test that hides the search bar for the control group
    • dReport the 150% lift as proof that search causes higher conversions
    Show the answer

    Answer: a · Match searchers to similar non-searchers using device, landing page, and prior sessions, then compare rates

    Matching on confounders like device and landing page isolates selection bias, which is why the example shows the true incremental lift drops to roughly 25% after propensity matching. Running a two-sample t-test on the raw rates assumes the groups are comparable and would incorrectly validate an inflated, non-causal difference.

    Read the full bite: How would you validate that search users are more likely to purchase?

20. Question 20 of 30

    A photography platform archives original images for future editing and serves compressed previews. Which strategy best preserves fidelity while optimizing delivery?

    • aStore originals as JPEG at maximum quality and serve previews as PNG to eliminate compression artifacts.
    • bUse JPEG at the highest quality setting for both archives and previews to maintain visual consistency.
    • cKeep originals as PNG and serve previews as PNG to preserve exact pixel data at every stage.
    • dStore originals in a lossless format and serve previews as lossy JPEG to reduce bandwidth.
    Show the answer

    Answer: d · Store originals in a lossless format and serve previews as lossy JPEG to reduce bandwidth.

    Lossless storage preserves bit-exact originals for future editing and avoids generational degradation, while lossy JPEG cuts preview sizes for faster web delivery. Option C is tempting because PNG guarantees exact pixels, but serving previews losslessly wastes bandwidth without perceptible quality gains over a well-compressed lossy image.

    Read the full bite: What is the difference between lossy and lossless image compression?

21. Question 21 of 30

    Which scenario would trigger an automated MLOps deployment but typically not a traditional DevOps pipeline?

    • aA scheduled rebuild of the serving application container
    • bProduction monitoring detecting drift in input data distributions
    • cA pull request merging changes to the model serving API
    • dA new GPU driver update pushed to the inference cluster
    Show the answer

    Answer: b · Production monitoring detecting drift in input data distributions

    The card states that MLOps deployments add triggers like data drift detection, unlike DevOps pipelines that react to code commits, dependency patches, or infrastructure changes. The GPU driver update is a tempting distractor because a common misconception is that MLOps is simply DevOps plus GPUs.

    Read the full bite: What are the primary differences between traditional DevOps and MLOps?

22. Question 22 of 30

    Which statement best describes how histogram equalization remaps grayscale intensities?

    • aIt counts pixels in each bin to detect whether the image is under- or over-exposed.
    • bIt linearly stretches the darkest and brightest pixels to fit the full 0-255 range.
    • cIt shifts all pixel values by a constant to center the histogram around the mid-tone.
    • dIt applies the cumulative distribution function to redistribute intensities toward a uniform histogram.
    Show the answer

    Answer: d · It applies the cumulative distribution function to redistribute intensities toward a uniform histogram.

    Histogram equalization uses the cumulative distribution function of the original histogram to remap intensities so the output approximates a uniform distribution, maximizing global contrast. The first option describes linear contrast stretching, which merely rescales the minimum and maximum values without considering the frequency of each intensity level.

    Read the full bite: Describe a grayscale histogram and its use in exposure and equalization

23. Question 23 of 30

    Beyond modeling long-range dependencies, what key training advantage does self-attention have over an LSTM?

    • aIt requires no positional information because order is built in
    • bIt eliminates the need for any learned weight matrices
    • cIt can process all sequence positions in parallel rather than step by step
    • dIt has lower computational complexity in sequence length than recurrence
    Show the answer

    Answer: c · It can process all sequence positions in parallel rather than step by step

    Self-attention computes all token interactions simultaneously, enabling parallel training that sequential LSTMs cannot match. It actually needs added positional encodings and is quadratic, not cheaper, in sequence length.

    Read the full bite: Self-Attention versus Recurrent Architectures

24. Question 24 of 30

    Why should batch product metadata and streaming click events typically use different physical storage formats?

    • aColumnar formats should be used for both to maximize compression and query performance across all workloads
    • bThe choice is purely based on cost, so the cheapest format should be selected regardless of data velocity or access pattern
    • cRow-oriented formats minimize write overhead for high-velocity events, while columnar formats improve projection and compression for batch analytics
    • dRow-oriented formats are preferred for both because they simplify schema evolution and reduce system complexity
    Show the answer

    Answer: c · Row-oriented formats minimize write overhead for high-velocity events, while columnar formats improve projection and compression for batch analytics

    Row-oriented formats minimize write overhead for streaming ingestion while columnar formats allow efficient projection and compression for batch training data. Choosing columnar for both is tempting but adds unnecessary write overhead to high-velocity events.

    Read the full bite: Design ingestion for clickstream and batch product metadata

25. Question 25 of 30

    When implementing a box blur, why is it important to write results into a separate destination buffer rather than updating the source image in place?

    • aIt lets the algorithm handle borders without explicit boundary checks
    • bIt ensures intermediate sums can be stored in a larger data type to avoid overflow
    • cIt prevents already-blurred pixel values from being reused in later neighborhood averages
    • dIt reduces memory usage compared to allocating a second image array
    Show the answer

    Answer: c · It prevents already-blurred pixel values from being reused in later neighborhood averages

    Using a separate destination buffer guarantees that every neighborhood average reads only original pixel values, not values that have already been blurred and would distort subsequent averages. The overflow issue in option B is addressed by using a larger type for the accumulator during the sum, not by allocating a second image buffer.

    Read the full bite: How would you implement a simple box blur on a grayscale image?

26. Question 26 of 30

    Why does ELT better support iterative ML experimentation than ETL?

    • aIt enforces strict schemas and data types before any data is loaded
    • bIt is primarily designed for structured legacy databases with stable columns
    • cIt transforms data on a secondary server before loading it into the warehouse
    • dIt allows repeated transformations of raw data within the warehouse without re-extraction
    Show the answer

    Answer: d · It allows repeated transformations of raw data within the warehouse without re-extraction

    ELT loads raw data into the target warehouse first, so data scientists can run and revise transformations repeatedly without rebuilding external pipelines or re-extracting source data. Option C describes ETL, which transforms data on a secondary server before loading and requires pipeline changes for new logic.

    Read the full bite: ETL vs ELT: when to prefer each for ML?

27. Question 27 of 30

    A grayscale image looks dull because most pixels are clustered between intensity 100 and 150. After histogram equalization, what has fundamentally changed about the pixel intensities?

    • aThe darkest pixel was mapped to 0 and the brightest to 255 with linear scaling between them.
    • bThe pixels were multiplied by a constant factor to brighten the entire image uniformly.
    • cA transfer function based on the cumulative intensity distribution was used as a lookup table to spread values across the full range.
    • dThe pixels were redistributed so that every intensity level from 0 to 255 contains an equal number of pixels.
    Show the answer

    Answer: c · A transfer function based on the cumulative intensity distribution was used as a lookup table to spread values across the full range.

    Histogram equalization computes the cumulative distribution function from the histogram, normalizes it to the maximum intensity, and uses it as a lookup table to remap pixels across the full range. Option A describes linear contrast stretching, which only scales the min and max values without considering the actual probability distribution of intensities.

    Read the full bite: What is an image histogram and how does histogram equalization improve contrast?

28. Question 28 of 30

    Which upstream strategy best prevents a categorical encoder from crashing when new values appear in training data?

    • aEnforce a locked schema that rejects batches with out-of-domain categories before encoding
    • bSilently drop unseen categories inside the encoder so training continues
    • cRely solely on an OOV fallback bucket in the encoder without upstream checks
    • dLet the encoder auto-update its vocabulary whenever a new category arrives
    Show the answer

    Answer: a · Enforce a locked schema that rejects batches with out-of-domain categories before encoding

    Enforcing a locked schema upstream acts as a hard gate that stops bad batches before they reach the encoder and forces explicit review for domain changes. Relying solely on an OOV bucket is wrong because the card treats it only as a last-resort safety net, not a primary data-quality strategy, and using it alone can mask upstream data bugs.

    Read the full bite: What data validation strategy prevents new categories from breaking your encoder?

29. Question 29 of 30

    What fundamentally distinguishes BERT's masked language modeling from GPT's causal language modeling?

    • aMLM is supervised with human labels while CLM is fully self-supervised
    • bMLM generates text autoregressively while CLM only classifies
    • cCLM uses bidirectional attention while MLM is strictly left-to-right
    • dMLM predicts masked tokens using bidirectional context; CLM predicts the next token from left context only
    Show the answer

    Answer: d · MLM predicts masked tokens using bidirectional context; CLM predicts the next token from left context only

    MLM reconstructs hidden tokens using context from both sides, suiting understanding, while CLM's left-only next-token prediction suits generation. Both are self-supervised, and it is CLM, not MLM, that generates autoregressively.

    Read the full bite: Causal versus Masked Language Modeling

30. Question 30 of 30

    Why does the Harris detector consider a corner a more reliable tracking feature than a point along a straight edge?

    • aA corner constrains position in two directions, while an edge point can slide along the edge
    • bEdges cannot be detected by image gradients
    • cCorners are always brighter than surrounding pixels
    • dCorners require no gradient computation to locate
    Show the answer

    Answer: a · A corner constrains position in two directions, while an edge point can slide along the edge

    At a corner, intensity changes in every shift direction, giving large eigenvalues and a precise 2D location. Along an edge only one direction constrains position, so the point slides (the aperture problem).

    Read the full bite: Harris corner detector and corner stability