Step 2: Collect and Upload the Dataset

Shoot the photos with the ESP32-CAM itself where you can. A phone camera produces sharper, better-lit images than the OV2640, and a model trained on those tends to disappoint on the real hardware - the sensor it meets in production is the noisy one.

Uploading

1. Go to Data acquisition in the left sidebar.
2. Click the Upload data icon.
3. Select your image files, or a folder.
4. Set Upload into category to Automatically split between training and testing.
5. Leave Label blank - bounding boxes are drawn in the next step.
6. Click Upload data and wait for the queue to clear.

Aim for an 80/20 train/test split, which the automatic option gives you. Keep the test set genuinely unseen - reusing training photos there produces a flattering accuracy number that tells you nothing.

Step 3: Label with Bounding Boxes

1. Open Data acquisition → Labeling queue.
2. Drag a box around each object in the image.
3. Type the class name, e.g. resistor. Reuse the exact same spelling every time - Resistor and resistor become two separate classes.
4. Click Save labels to advance to the next image.

Edge Impulse tracks objects between similar frames, so after the first few images the boxes are often pre-drawn and you are only correcting them. Label every instance in the frame - an unlabelled object teaches the model that this pattern is background, which is actively harmful.

Step 4: Design the Impulse

1. Go to Impulse design → Create impulse.
2. Set image width and height to 96 × 96 and resize mode to Fit shortest axis.
3. Add an Image processing block.
4. Add an Object Detection (Images) learning block.
5. Click Save Impulse.
6. Open the Image block, set colour depth to Grayscale, and click Save parameters → Generate features.

96 × 96 grayscale is not a stylistic choice. The ESP32 has roughly 4 MB of PSRAM and no vector unit. RGB at 160 × 160 will either fail to allocate its tensor arena at boot or run several seconds per frame. Grayscale cuts input by two-thirds for little accuracy loss on shape-driven objects.

Step 5: Train the Model

1. Open the Object detection block.
2. Choose FOMO (MobileNetV2 0.35) as the model architecture.
3. Start with 60 training cycles and a learning rate of 0.001.
4. Click Start training and wait - a few minutes for a small dataset.

Why FOMO: conventional detectors like YOLO or MobileNet-SSD will not fit in the ESP32 flash budget. FOMO predicts object centroids on a coarse grid instead of full boxes, which is around 30× smaller and fast enough for real-time video on this hardware. The trade-off is that it gives you object positions and counts, not precise box dimensions - usually fine for "how many, and roughly where".

Aim for F1 above 0.8. If it is lower, the fix is almost always more varied training data - different lighting, angles and backgrounds - rather than more training cycles. Past a point, extra epochs only memorise the training set.

Step 6: Test Before You Flash

1. Go to Model testing and click Classify all.
2. Review accuracy against the held-out test set.
3. Use Live classification to try a single new image.

Do this before touching hardware. Debugging a weak model through a serial monitor is slow and misleading - if accuracy is poor here, it will be poor on the board, and no amount of firmware tinkering will fix it.

Step 7: Export as an Arduino Library

1. Go to Deployment.
2. Search for and select Arduino library.
3. Leave optimisations on Quantized (int8) - the float32 build is markedly slower and larger.
4. Click Build. A .zip downloads when it finishes.

In the Arduino IDE, go to Sketch → Include Library → Add .ZIP Library and select the file you just downloaded. Do not unzip it first.

Step 8: Open and Modify the Example

Open File → Examples → <your-project-name>_inferencing → esp32 → esp32_camera. The sketch compiles almost as-is; one edit is mandatory.

The remaining settings are optional. Leave them alone on a first run and come back if performance disappoints.

Step 9: Board Settings

In Tools, set the following. The partition scheme is the one people miss.

• Board: AI Thinker ESP32-CAM
• Partition Scheme: Huge APP (3MB No OTA/1MB SPIFFS)
• PSRAM: Enabled
• Upload Speed: 115200 - drop to 57600 if uploads fail intermittently
• Flash Frequency: 80 MHz

If the board does not appear: add https://espressif.github.io/arduino-esp32/package_esp32_index.json under File → Preferences → Additional Board Manager URLs, then install esp32 by Espressif Systems from the Boards Manager.

On the partition scheme: the default leaves roughly 1.2 MB for the application, and a FOMO model plus the camera driver exceeds that. The symptom is a compile-time "text section exceeds available space in board" error, which reads like a code problem and is not one.

Step 10: Wire and Flash

Wire the FTDI adapter to the ESP32-CAM:

• 5V → 5V
• GND → GND
• TX (FTDI) → U0R (board)
• RX (FTDI) → U0T (board)
• IO0 → GND - only while flashing

Note that TX goes to U0R and RX goes to U0T. The lines cross - wiring them straight through is the single most common reason an upload never starts.

Flashing Sequence

1. Connect IO0 to GND to enter bootloader mode.
2. Plug in the FTDI adapter, then tap the RST button on the board.
3. Click Upload.
4. When "Connecting…" appears, the board is listening. If it stalls on dots, tap RST again.
5. After upload completes, disconnect IO0 from GND.
6. Tap RST once more to boot the new firmware.

Step 11: Test It

1. Open Tools → Serial Monitor.
2. Set the baud rate to 115200.
3. Tap RST.
4. Hold a trained object in front of the lens.

Working output looks like this:

The numbers in brackets are grid coordinates in the model input space, not pixels in the captured frame. Classification time around 180 ms is normal for FOMO at 96 × 96, giving roughly 5 frames per second.

Troubleshooting

• Camera probe failed (0x105): the ribbon cable is loose, or more than one CAMERA_MODEL_* is defined. Reseat the connector and re-check Step 8.
• Brownout detector was triggered: insufficient power. Use the 5V pin, not 3.3V, and prefer a separate supply - many FTDI adapters cannot source the camera’s peak current.
• Guru Meditation Error / boot loop: PSRAM is disabled, or the partition scheme is wrong. Re-check Step 9.
• Failed to connect / stuck on dots: IO0 is not grounded, RX and TX are not crossed, or RST was not tapped after entering bootloader mode.
• Detects nothing on hardware, but tests well in the Studio: the training images were too clean. Re-shoot part of the dataset with the ESP32-CAM itself, under the lighting it will actually run in.

Where to Go Next

• Stream annotated frames over Wi-Fi with the esp32_camera_web example in the same library.
• Drive a GPIO on detection - a relay, a buzzer, a servo - straight from the inference callback.
• Retrain with images captured by the deployed board to close the gap between test and field accuracy.