Interface choice is usually made last, after resolution and frame rate have been settled - and that is exactly backwards. The interface determines how far the camera can sit from the PC, how many cameras you can run from one host, and whether you will be able to add a seventh station to a line that already has six. It is the hardest decision to reverse.
Do the bandwidth arithmetic first
Raw data rate is width times height times bit depth times frames per second. A 5 MP mono camera outputting 8-bit pixels at 24 fps needs 5.0 x 1024 x 1024 x 8 x 24 bits per second, roughly 1.0 Gbps - almost exactly the usable payload of GigE Vision, which is why so many 5 MP GigE cameras are quoted at 23 to 24 fps and not at 60. The same camera on USB3.0 has roughly 3.2 Gbps available and can run far faster, or run at the same rate with headroom for a second camera.
Scale that up: a 20 MP camera such as the MV-CS200-10GM on a single GigE link is limited to single-digit frame rates. That is often perfectly acceptable for static presentation inspection, and completely unacceptable for a moving web. The same 20 MP sensor with a USB3.0 interface, as in the MV-CS200-10UM, is the right pick when you need 20 MP and can accept a 3 to 5 metre cable run.
GigE Vision: the default for a reason
Poe-capable GigE gives you a single cable per camera for data and power, runs 100 metres without a repeater, and scales cleanly through an industrial switch. Multi-camera systems are straightforward: a 12-port managed switch with PoE will happily feed a dozen 5 MP cameras from one host, provided the host's NIC and the switch uplink have the aggregate bandwidth. GigE is also the least driver-hostile option on Windows - jumbo frames and a dedicated NIC are usually the only tuning needed.
Its limits are equally clear. Aggregating four cameras onto one host means aggregating their traffic too, and a 1 Gbps uplink saturates at about four 5 MP cameras running at full rate. Packet loss under load shows up as torn frames, which is a far worse failure mode than a simple frame-rate shortfall.
USB3 Vision: cheap speed, short reach
USB3.0 delivers roughly 3.2 Gbps of usable payload at a fraction of the cost of 10GigE, which makes it the value choice for single-camera stations. Passive cables are reliable to about 3 metres; beyond that you need active optical cables or a powered hub, and those bring their own power and enumeration quirks. Power delivery over USB3 is limited, so higher-power cameras may need a separate supply.
The practical rule: one camera, short cable run, frame rate matters - USB3. Three or more cameras, long runs, or a switch-based topology - GigE.
10GigE and CoaXPress: for the top of the range
Once you move past roughly 25 MP at speed, or into multi-line line-scan configurations such as the MV-CL042-91GM where the sensor is streaming continuously, 1 Gbps stops being enough and 10GigE or CoaXPress becomes the only option. A 250 MP large-format camera like the MV-CH250-90TM-M58S-NF is not going to be fed by GigE at any useful rate. Budget for the NIC, the switch and the cabling as part of the camera cost - 10GigE is not a drop-in upgrade to a GigE station.
A worked example
A six-station bottling line, each station running a 5 MP mono camera at 20 fps in 8-bit. Per-station raw load is about 0.85 Gbps. Six stations is 5.1 Gbps, which cannot share a single 1 Gbps uplink. Two reasonable answers: three dual-port GigE NICs in the host with separate subnets, or one 10GigE switch with 10GigE uplinks and per-port GigE downlinks. The second is tidier and easier to expand; the first is cheaper and adequate if you will never add station seven.
The takeaway: decide bandwidth from the sensor and the cycle time, decide topology from the number of cameras and the cable distance, and only then pick the product family.